DEFENDIFY: defense amplified with transfer learning for obfuscated malware framework

CastilloCamargoetal. Cybersecurity (2025) 8:97 https://doi.org/10.1186/s42400-025-00396-z RESEARCH Open Access © The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Cybersecurity DEFENDIFY: defense amplified withtransfer learning forobfuscated malware framework Rodrigo Castillo Camargo1, Juan Murcia Nieto1, Nicolás Rojas4, Daniel Díaz‑López1* , Santiago Alférez3, Angel Luis Perales Gómez2, Pantaleone Nespoli2, Félix Gómez Mármol2 and Umit Karabiyik5 Abstract The existence of malicious software (malware) represents a potential threat to users who connect to a large set of services provided by multiple providers. Such malware is capable of stealing, spying on, encrypting data from users, and spreading, provoking impacts that are beyond a single citizen’s device and reaching critical information systems. To detect malware families, Machine Learning and Deep Learning techniques have been employed recently, demon‑ strating promising results. However, these techniques lack in detecting more advanced malware that employs obfus‑ cation techniques. In this paper, we present DEFENDIFY, a novel framework, empowered by Computer Vision, Deep Learning, and Transfer Learning techniques, that is able to detect completely obfuscated malware with high perfor‑ mance in terms of accuracy and computational consumption. DEFENDIFY comprises three modules: Dataset Crea‑ tion, Binary Obfuscation, and Model Generation. These modules work together to detect both obfuscated and non‑ obfuscated malware. The core module, i.e., the Model Generation, employs an entropy tester that determines whether a sample is obfuscated or not. Then, a Deep Learning model powered by Transfer Learning is employed to determine if it is malware or goodware. We validated our framework using real data gathered from malware repositories and legitimate software. The proposed framework was configured to test four Convolutional Neural Network architectures: ResNet18, ResNet34, EfficientNetB3, and EfficientNetV2S. Among them, the ResNet18 archi‑ tecture obtained the best performance in detecting both non‑obfuscated and obfuscated samples with an F1‑score of 99.34% and 97.5%, respectively. Keywords Malware detection, Malware obfuscation, Computer vision, Transfer learning, Deep learning, Networking system of artificial intelligence Introduction With the rise of new technologies such as the Internet of Things (IoT)(Glaroudis etal. 2020) and Big Data (BD) (Iaksch et al. 2021), cities have become increasingly automated, thus embracing the smart cities paradigm(Ristvej etal. 2020). This has undoubtedly provided a clear advantage for all citizens, enabling access to unimaginable services just ten years ago. However, there arises an urgent need for safeguarding the digital backbone. As we immerse ourselves in the transformative potential of IoT and BD within city infrastructures, the vulnerability to cyber threats becomes increasingly pronounced, and therefore, new dangers loom over the inhabitants of a smart city. *Correspondence: Daniel Díaz‑López [email protected] 1 School of Engineering, Science and Technology, Universidad del Rosario, Kr 6 12c‑16, Bogotá 111711, Cundinamarca, Colombia 2 Faculty of Computer Science, University of Murcia, Campus de Espinardo, 30100 Murcia, Murcia, Spain 3 Department of Mathematics, Barcelona East Engineering School, Polytechnic University of Catalonia, 610101 Barcelona, Catalonia, Spain 4 School of Engineering, Pontifical Xavierian University, Bogotá 610101, Cundinamarca, Colombia 5 Department of Computer and Information Technology, Purdue University, West Lafayette, IN 610101, USA Page 2 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 One of the critical threads in smart cities is cybersecurity and, especially, the proliferation of malware(Ma 2021). Specifically, this term is used to refer to malicious programs designed to perform harmful actions on user’s devices(Useche-Pelaez etal. 2018). These actions may include spying on users, stealing sensitive information, damaging the victim’s device, or executing other malicious actions, to cite some examples(Martínez Martínez etal. 2021). Different smart city services could be affected by a malicious campaign, e.g. traffic management, pollution monitoring, water quality control, public safety support, etc. Furthermore, there are also organized campaigns designed to spread malware and target different victims, as dissemination and exploitation of malware have become a profitable business(Hakon etal. 2020). As new forms of malware are constantly being created and spread, the anti threat industry has sought to mitigate the risk of a machine being infected with a malicious program through various malware detection techniques. These techniques include the use of different types of signatures, such as host-based, malware, and network signatures(Pelaez etal. 2018; Nespoli etal. 2019), or ML algorithms such as support vector machine (SVM) and Random Forest, achieving promising results(Joshi etal. 2018). Due to the advancement in antivirus industry, malware evasion techniques have also been developed to reduce the effectiveness of the proposed detection methods. Among others, a well-known evasion technique leverages the use of encoders or obfuscators, which are coding functions that change the appearance of programs, thus making them unrecognizable to antiviruses and allowing attackers to bypass protection layers when delivering malware(Kim etal. 2018). Nevertheless, it is important to note that not all encoded or obfuscated binaries are malicious programs. There is also well-intentioned software, commonly referred to as goodware, that is encoded by its developers to protect it from industrial espionage or tampering (Toffalini et al. 2019). In this paper, the words encoded/obfuscated and non-encoded/non-obfuscated will be used indistinctly. In this scenario, some studies proposed the use of artificial intelligence (AI), and especially machine learning (ML) and deep learning (DL) techniques to detect malware. Among these techniques, DL models focused on the computer vision (CV) field are gaining prominence due to their ability to automatically extract features that can capture subtle patterns in obfuscated code that might be invisible to traditional analysis methods.. The procedure is to consider malware as images and use image classification to detect malware by identifying key features of the image(Pereira-Kohatsu etal. 2019; Xue etal. 2019; Bensaoud etal. 2020). The most critical limitation of these approaches and those based on traditional ML algorithms or signatures is that they do not allow one to detect completely obfuscated malware. Furthermore, the procedures to train these models are highly expensive regarding the computation time since they need to train an entire DL model from scratch, preventing the wide adoption of these systems in current technological scenarios or in future smart city architectures. One technique that can improve the result obtained in this field is the transfer learning (TL)(Weiss etal. 2016), which consists on apply the knowledge acquired in one domain such as CV on another new domain such as malware classification. By transferring learned representations from pre-trained models on image datasets, it is possible to adapt these insights to the specific task of malware detection. This strategy not only reduces the training time required but also significantly enhances detection performance, making it a more efficient and effective solution for identifying complex, hidden threats. The major motivation for our research stems from three critical challenges in current malware detection approaches. First, there is a significant gap in effectively detecting completely obfuscated malware using static analysis techniques, which is essential for identifying threats before execution. Second, the substantial computational resources required for training deep learning models from scratch make many current solutions impractical for widespread deployment. Third, there is a lack of comprehensive frameworks that integrate different components necessary for effective detection of both obfuscated and non-obfuscated malware in real-world scenarios. Additionally, another outstanding component in detecting malware samples is the entropy tester (BatErdene et al. 2017). In particular, the entropy-based obfuscation tester can be described as a fundamental component as part of the static analysis process, allowing for the classification of both obfuscated and non-obfuscated binaries without needing dynamic analysis. Last but not least, a critical challenge for AI-powered malware detection is the lack of diverse and comprehensive datasets (Allix et al. 2015), particularly when it comes to handling encoded or obfuscated samples. Without access to a variety of datasets, including nonencoded, XOR-encoded, and Shikata Ga Nai-encoded goodware and malware, AI models are unable to fully learn the complexities of both obfuscated and nonobfuscated threats. This gap severely limits the ability of existing solutions to generalize across different encoding techniques used by malware authors to evade detection. These challenges underscore the urgent need for a novel, integrated approach that can effectively detect obfuscated malware while remaining computationally Page 3 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 efficient. Our research is driven by the vision of creating a practical and deployable framework that addresses these limitations by leveraging the strengths of both computer vision and transfer learning, while incorporating specialized components to handle different types of malware, whether obfuscated or not. In light of this, this paper proposes DEFENDIFY, a novel framework empowered by CV, DL, and TL(Ribani and Marengoni 2019), to detect obfuscated malware samples. DEFENDIFY may be applied to protect user’s devices, or particular smart citizens’ devices under a smart city scenario. In this sense, TL refers to the use of knowledge acquired from one task to improve performance on a related but different task without training again the whole DL model. That is, the models are pretrained on various image datasets, and the knowledge acquired is then used as baseline to classify obfuscated and non-obfuscated malware, avoiding a potentially expensive training phase. With this in mind, and in order to make use of the knowledge obtained from the pretrained CV artifacts, the ingested binaries must be transformed into images compatible with the original network architecture. The reason for this is that, although TL is a powerful learning tool, it has some restrictions to use the same type of data format with which the “transferred” knowledge was generated. Specifically, DEFENDIFY consists of three modules: Dataset Creation, Binary Obfuscation, and Model Generation. These modules work together to detect completely obfuscated malware with a high performance in terms of accuracy and computational consumption. The use of CV techniques and convolutional neural networks (CNN) is particularly well-suited for image classification tasks and has shown promising results in various CV applications. In our proposal, the binary code of obfuscated malware is converted into image, aiming to provide a new perspective for malware detection, which can overcome the limitations of traditional detection methods and improve the detection of obfuscated malware. To summarize, the specific contributions of this paper are as follows: 1. A novel and generic framework powered by DL and TL technologies that consists of three modules to detect both obfuscated and non-obfuscated malware. To be specific, the framework incorporates an entropy tester to differentiate obfuscated and nonobfuscated samples, and two DL models powered by TL that classify goodware and malware. 2. The construction of three datasets of non-encoded, XOR-encoded and Shikata Ga Nai-encoded goodware and malware samples. These datasets were generated by the framework itself thanks to the Dataset Creation module. In later steps, these datasets were converted into images by the Binary Obfuscation module using a specific codec. 3. The validation of DEFENDIFY in terms of evaluation performance and computational consumption. DEFENDIFY was configured to test four CNN architectures: ResNet18, ResNet34, EfficientNetB3, and EfficientNetV2S. In terms of evaluation performance, the best architecture was RestNet18 that reached an F1-score of 99.34% and 97.24% for non-obfuscated and obfuscated malware detection, respectively. In terms of computational consumption, the most efficient architecture was RestNet18 followed by RestNet34, EfficientNetB3, and finally EfficientNetV2S. 4. A functional full-fledged prototype of DEFENDIFY, offered as an IA-enabled smart service for the detection of advanced malware, which is accessible to the scientific community who desire to test it and validate its practicality and easiness-to-use. The prototype is accessible at: https:// huggi ngface. co/ spaces/ esab/ malwa re_ detec tion The remainder of the manuscript is structured as follows. “Leveraging computer vision for cybersecurity” section contains a overview of how CV may support cybersecurity challenges. “State of the art” section compares the main works proposed in this field, highlighting its contributions and techniques. Then, “DEFENDIFY framework” section describes the details of the DEFENDIFY framework. In “Experiments”, the experiments that were carried out are shown in detail. “Discussion” section introduce a brief discussion about the work presented in this paper. At last, “Conclusions and future work” section concludes the work, summarizing the results and showing the future directions of this research. Leveraging computer vision forcybersecurity Computer vision (CV) has emerged as a transformative technology in the field of cybersecurity due to its ability to analyze and interpret complex data patterns. By converting data into visual representations, CV techniques offer innovative approaches for identifying and mitigating cyber threats, particularly in scenarios involving malware detection and analysis. One of the key advantages of CV is its applicability to static malware analysis. By transforming binary files into grayscale images, convolutional neural networks (CNNs) can be employed to detect patterns indicative of malicious code. These models excel at recognizing features associated with obfuscation techniques or specific malware families, enabling high-accuracy classification even in challenging cases (Li et al. 2022). For example, CV-powered classifiers can analyze code structures Page 4 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 or entropy patterns to identify hidden malicious binaries, supporting static code analysis of logs and network frames collected from information systems or network devices. CV can also be utilized for dynamic threat detection. In network security, CNNs can analyze traffic patterns to identify anomalies, such as unusual port activity or data exfiltration attempts. These methods enhance the ability to detect both known and zero-day attacks, contributing to the protection of critical systems(Li etal. 2022). Additionally, CV-based solutions can support phishing detection by analyzing visual and structural features of email campaigns. CNNs trained on phishing patterns can filter suspicious content, providing a proactive layer of defense against social engineering attacks. This is particularly relevant for organizations exposed to high volumes of communication data, where manual analysis would be infeasible. Examples of similar approaches include applications by Mimecast CyberGraph and Cofense Cyberfish(Mimecast 2025; Cyberfish 2025). Another application of CV in cybersecurity is its role in countering adversarial attacks. Attackers often attempt to bypass machine learning models by introducing subtle perturbations in input data. CV techniques can enhance model robustness by employing advanced architectures and data augmentation strategies, reducing the effectiveness of adversarial mechanisms(Xi 2020). Finally, CV is instrumental in monitoring physical behaviors that may indicate malicious intent. Techniques such as micro-expression analysis, posture recognition, or facial stress detection can be used to identify deceptive activities, which are critical in scenarios like border security or physical access control(Rehman etal. 2022). By integrating CV into cybersecurity workflows, researchers and practitioners can develop robust defenses against evolving cyber threats. Its ability to extract meaningful insights from complex data makes it a valuable tool in advancing malware detection and enhancing overall cybersecurity, as will be seen in the paper at hand through the proposal of DEFENDIFY framework. State oftheart In this section, we review the available literature in the field of malware detection. The traditional approach to detect them can be classified into signature-based, anomaly-based and specification based(Talukder and Talukder 2020). It is worth mentioning that with the arrival of ML and DL techniques, this field has experienced a huge change and many of the proposed works use the anomaly detection (AD) paradigm (Sahin and Bahtiyar 2020; El Merabet and Hajraoui 2019). One of the advantages of this approaches is that they can detect new types of malware without collecting its signatures. Due to the widespread use of x86 architectures, most work on malware detection focuses on this architecture. Thus, the first question that needs to be answered is about the features used to feed ML and DL models for malware detection in such architecture. These features need to be sufficiently representative of the samples to discriminate between malware and goodware. Until now, there were four main groups of features to use in malware detection: Opcode sequence, portable executable (PE) header, Strings, and API sequence (Guo and Fan 2019). Besides, the study presented in(Xue etal. 2019) shows a taxonomy of ML methods for Binary Code Analysis, including an exhaustive list of the features used in malware detection. In this context, many of the ML techniques currently applied for malware detection are based on these features. For example, Rezaei etal. (2021) propose a novel method that combines DL and ML techniques to learn different embedding representations of both malware and goodware. In particular, the raw bytes of PE header are embedded into the dense neural network (DNN), and the output is passed to a k-means model. The authors highlight the low computational overhead of its solution due to the lightweight network used and the low time required to extract bytes from the PE header. In the same context, several ML-based approaches have been proposed for malware detection by leveraging features extracted from PE headers. For instance, Hussain etal. (2022) introduce a detection system that applies various ML models, including Random Forest, support vector machine (SVM), and Gradient Boosting, to classify executables as clean or malicious based on PE header features. The authors conduct a comparative analysis of these models, highlighting that Random Forest achieves the highest accuracy (99.44%), making it a promising candidate for real-world malware detection applications. In terms of opcode sequence approaches, several works have been proposed. For example, Lee etal. (2023) proposed a method that extracts fixed-length and lowdimensional features from opcode category information to distinguish between benign and malicious application binaries. The extracted features are evaluated using multiple ML models, including SVM, Decision Tree, and Random Forest, achieving an accuracy of over 98% for both malware detection and classification. The authors highlight the robustness of their approach, demonstrating its effectiveness in identifying different malware families. In Kakisim etal. (2022), the authors proposed Sequential Opcode Embedding-based Malware Detection (SOEMD), a method that captures common malicious patterns using a Random Walk approach for edge and node selection. By constructing a low-dimensional vector space with opcode embeddings, SOEMD Page 5 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 enhances detection efficiency. The model tested in the architecture were K-Nearest Neighbour (k-NN), CNN and LSTM. Experimental results show that the proposed method outperforms baseline approaches, achieving a 100% malware detection rate. In terms of feature comparison, Balram etal. (2019) presented a detailed study where different ML techniques are tested using PE header and Strings. To be specific, the ML models tested were SVM, Linear Regression (LR), Random Forest (RF), XGBoost. Besides, based on these models, the authors also tested two ensemble models : LR/XGBoost and LR/RF/Naïve Bayes. The authors concluded that models that use string-based features outperform models that use PE-based features. The model that achieved the best performance was the ensemble of LR/ XGBoost with an accuracy of 0.980 for string-based features and 0.915 for PE-based features. Although both classical ML models and traditional features-based DL methods achieve good results, new approaches based on DL techniques are explored in the literature. In particular, convert binary files into images and using them as input of a CNN is achieving good results. In this context, Shaukat etal. (2023) proposed a deep learning-based method that visualizes PE files as colored images and extracts deep features using a finetuned deep learning model. These features are then classified using an SVM, eliminating the need for extensive feature engineering. The proposed method achieves 99.06% accuracy on the Malimg dataset and demonstrates superior performance over state-of-the-art approaches, with an average accuracy improvement of 16.56%. This new approaches based on DL models allow the detection of obfuscated malware. This is clearly show by Mercaldo etal. (2023) that proposed a method that converts system call traces from legitimate, malicious, and obfuscated Android applications into images for classification using a CNN. Their experiments demonstrate the resilience of deep learning models in detecting obfuscated malware, highlighting the effectiveness of dynamic analysis combined with CNN-based classification. Additionally, the authors employ explainability techniques to analyze model predictions, ensuring interpretability and robustness. Another example is exposed by Han etal. (2024), where they proposed a deep learning-based method that leverages depth-wise CNN with a spatial attention mechanism to classify malware obfuscated by virtual machine (VM) code. Using a dataset generated with VMProtect, the proposed model is trained on real-world obfuscated malware samples. Experimental results show that the approach achieves nearly 100% accuracy on regular malware classification and 93.55% on obfuscated malware, demonstrating its effectiveness in handling complex obfuscation techniques. Ravi etal. (2022) propose a Multi-View attention-based DL framework for malware detection, leveraging features from PE headers, imports, images, and API calls. Their approach outperforms nonattention-based ML models, achieving 95% accuracy. Additionally, they evaluate malware detection using grayscale images from binary files, obtaining 98% accuracy on a Windows dataset and 97% on an Android malware dataset. Finally, Conti etal. (2022) explore DL models for detecting and classifying obfuscation in Android applications. They evaluate classical ML methods alongside NLP and image recognition techniques on binaries obfuscated using four strategies: Trivial (T), String Encryption (SE), Reflection (R), and Class Encryption (CE). Among the tested models (SVM, DNN, and CNN), CNNs using RGB images achieved the highest F-scores, with a peak of 0.994 for CE. A hybrid approach combining multiple models further improved performance, reaching F-scores of 0.972 (T), 0.991 (SE), 0.980 (R), and 0.998 (CE). Table1 compares the solutions discussed in the previous paragraphs together with the solution that we propose. From this comparison, it is evident that our approach is the only one offering a comprehensive framework for malware detection. Specifically, our contribution introduces several key innovations that distinguish it from prior work: • Use of Transfer Learning (TL) for obfuscated malware detection: Our method applies TL(Ribani and Marengoni 2019) to address the challenge of detecting fully obfuscated malware. This approach enables us to transfer learned representations from pretrained models on image datasets to the specific task of malware detection, thereby reducing training time and improving overall performance. • Integration of an entropy-based obfuscation tester: Our framework is unique in adopting a static analysis approach that classifies both obfuscated and nonobfuscated binaries without requiring execution of the binaries, thus mitigating the risks associated with dynamic analysis. Prior to feeding the samples into the CNN model, we perform an entropy test to distinguish between obfuscated and non-obfuscated samples. This step is novel, as no previous work integrates an obfuscation detection module as part of the classification pipeline. • Detection of fully obfuscated malware using realworld obfuscators: While the works proposed byConti et al. (2022) and Han etal. (2024) bears some resemblance to our method, there are two major distinctions: i) their solution does not present a complete framework, and ii) they do not leverage TL techniques. Additionally, unlike previous works, includingConti etal. (2022), our approach is the first Page 6 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 to validate the detection of fully obfuscated malware using obfuscators that are widely employed by realworld attackers. Specifically, our solution demonstrates robustness against Shikata Ga Nai, an obfuscation tool frequently used in actual attack scenarios, which has not been addressed in prior studies. • Some related works use TL in malware detection labors but for different proposals. Jian etal. Jiang et al. (2020) proposed a TL-based solution that aims to detect variations between the representation of an original obfuscated malware and some of its variations, considering that both the original and the obfuscated malware have an image representation. Marastoni etal.Marastoni etal. (2021) proposed a solution that intentionally applies code transformations to malware to build an augmented malware dataset, which is then used to construct a TL-based model to differentiate between obfuscated and no-obfuscated malware samples. Thus, these last related works are focused on classifying representations of malware, instead of a full-fledged solution that not only detects obfuscation but also detects if the sample is or is not malicious as DEFENDIFY does. In summary, our proposal is the first to offer a fully integrated framework for malware detection that combines TL, a static analysis approach, and an entropybased obfuscation tester, ensuring robust performance even against fully obfuscated malware, as validated using real-world obfuscation tools. Table 1 Comparison of different related works that propose malware detection using ML/DL Work Can detect completely obfuscate malware by means of static analysis? Dataset features ML/DL model Year Balram et al. (2019)✕Strings PE header SVM LR RF XGBoost Ensemble (LR/XGboost) Ensemble (LR/RF/NB) 2019 Rezaei et al. (2021)✕Raw bytes of PE header DNN K‑Means 2021 Ravi et al. (2022)✕PE header PE imports API calls Bytes of PE header as images attention‑based CNN, DNN, and LSTM 2022 Conti et al. (2022)✓Binaries as gray‑scale/RGB images Features extracted from the strings Natural language‑based features SVM DNN CNN 2022 Hussain et al. (2022)✕PE Header RF SVM Gradient Boosting 2022 Kakisim et al. (2022)✕OP code RF K‑NN CNN LSTM 2022 Lee et al. (2023)✕OP code RF Decision Tree SVM 2023 Shaukat et al. (2023)✕Binaries as RGB images CNN as feature extractor and SVM 2023 Mercaldo et al. (2023)✕system calls traces encoded as images CNN 2023 Shaukat et al. (2023)✕ Binaries as RGB images CNN as feature extractor and SVM 2023 Han et al. (2024)✓Binaries as RGB images CNN with attention mechanism 2024 DEFENDIFY ✓Binaries as grey‑scale images CV architectures powered by TL 2025 Page 7 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 DEFENDIFY framework In this section, the components and modules that comprise the DEFENDIFY framework are detailed. As previously mentioned, the proposed framework utilizes DL and TL technologies to detect obfuscated malware. DEFENDIFY consists of three modules: Dataset Creation (“Dataset creation” section), Binary Obfuscation (“Binary obfuscationy” section), and Model Generation (“Model generation” section). Each of these modules is further divided into various components that work in coordination to efficiently detect obfuscated malware. A highlevel description of these modules and components can be observed in Fig.1. Dataset creation In order to perform its detection activities, DEFENDIFY needs to recollect and store a large number of samples as input. To achieve this, the Dataset Creation module comprises a Malware Selection and a Goodware Selection component. The first is responsible for gathering malware binaries, while the second is in charge of collecting goodware binaries, as detailed next. Although the current section describes the generic procedure for the dataset composition, a description the datasets used in our experiments which impact directly the obtained performance metrics is detailed in “Datasets used in the experiments” section. Malware selection The Malware Selection component is in charge of one of the most important actions in DEFENDIFY, as it collect, filter and select the final set of malicious binaries that will be used in the next training steps. Even if there are so many repositories where malware may be collected, it is important to consider key features as the volume of samples, diversity in the types of malware, trust in the source, integrity of the samples, among others.Thus, there is different malware repositories (e.g., Malware The Zoo,1 vxvault,2 vxunderground3 or VirusShare4) which may be a good sources to be consumed by the Malware Selection component. From the previously mentioned repositories, VirusShare is particularly interesting as it is an open repository of malware samples that is intended to provide security researchers with samples of real malicious code. This repository was created to provide access to samples of live malicious code with the aim of helping researchers, forensic analysts, and similar professionals, however, to prevent the spread of the samples distributed in this repository, the access to those samples is delivered under individual request. As we will see in “Datasets used in the experiments” section, a dataset of 15821 malware samples was employed in our experiments. Goodware selection The Goodware Selection component provides DEFENDIFY the capacity to collect, filter and select the final set of non-suspicious binaries which will be used in the next training steps. There is a big diversity of goodware available online, downloadable through the original software publisher or third parties, addressed for different purposes, and types of public. Regarding the purpose, there are many e.g. libraries and handles used by operating Fig. 1 High‑level view of DEFENDIFY, the proposed malware detection framework 1 https:// github. com/ ytisf/ theZoo. 2 http:// vxvau lt. net/ ViriL ist. php. 3 https:// vxunder ground. org/. 4 https:// virus share. com. Page 8 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 systems, end user utilities, server software, middleware, etc. Thus, acquisition of goodware samples may be achieved through from the most naive approach based on a manual download of software typically used by users, until more automated ways that collect large amounts of goodware using application stores such as the one available at Microsoft Store.5 One interesting way of composing a proper goodware dataset, usable in DEFENDIFY, is including binaries that are found typically in an operating system, i.e. system files of Windows 10/11, and also including binaries associated to end user utilities, i.e. office applications. As we will see in “Datasets used in the experiments” section, a dataset of 15628 goodware samples was employed in our experiments. Binary obfuscation Once the malware and goodware samples have been recollected, the subsequent step entails the obfuscation of the samples. To this extent, the Binary Obfuscation module is responsible for obfuscating the samples gathered by the Dataset Creation module. In particular, this module comprises the Encoders Selection, Obfuscator, and Data Processing components. Encoders selection This component is in charge of selecting the encoders used to obfuscate both malware and goodware samples. In this regard, note that obfuscation may be achieved with different types of encoders, mainly metamorphic and polymorphic ones. Among the metamorphic ones, the block-based XOR encoder (Ceschin et al. 2021) is one of the most popular for its simplicity, because it reuses the idea of a mathematical XOR operation to perform transposition encoding, defeating in that way, the regular anti-malware mechanism based on signatures. This encoder uses the property of an XOR operation, i.e., A⊕B⊕B=A , and processes the binary per block. Thus, the steps to implement block-based XOR are resumed next: 1. Define a key kj to be used in the current block encoding. 2. Read and identify each character ci of the software shell code. 3. XOR each ci using kj . One of the goals of an encoder is to obfuscate the key shellcode operations existing in a binary. However, as XOR encoding is metamorphic, it causes known suspicious shellcode instructions, such as the corresponding to exec/bin/bash or /tcp/ ip port, become encoded to well-known outputs. From an offensive perspective, a well-known output represents a bad signature because it can be easily identified by anti-malware solutions. Metamorphic encoders evolved to polymorphic ones, meaning that two equal files received as input by the encoder will not produce the same output file, being Shikata Ga Nai one of the most representative and used polymorphic encoders(Farley and Wang 2014). Shikata Ga Nai is a polymorphic XOR additive feedback encoder that uses metamorphic techniques, e.g., reordering and substitution, in conjunction with a chained self-modifying key to produce a different output each time it is applied, bypassing detection mechanisms based on signature recognition(Graham 2021). Given that Shikata Ga Nai uses a chained self-modifying key through additive feedback, in case the input to be decoded or the keys are incorrect at any iteration, all subsequent outputs will be incorrect. The steps to implement Shikata Ga Nai are resumed as follows: 1. Define a key kj 2. Get the value of the Extended Instruction Pointer EIP register using Floating-Point Unit (FPU) instructions. 3. Enter in a loop that runs over a new instruction EIP+ 0xF. (a) Replace instruction at EIP+ 0xF applying an XOR operation between EIP+ 0xF and kj . (b) Change the key kj by adding it with the result of the previously modified instruction at EIP0xF. Although DEFENDIFY is generic enough to select any encoder in the literature, this work is focused on XOR and Shikata Ga Nai encoders. The reason for this decision was to cover both metamorphic and polymorphic encoders, and because they are encoders employed commonlyBurita and Le (2021). Obfuscator Once goodware and malware samples are collected by the Dataset Creation module, these samples must be obfuscated for the encoders selected for DEFENDIFY: XOR and Shikata Ga Nai. There are different tools to execute software obfuscation, being the Metasploit Framework (MF)6 one of the most recognized ones. MF is an opensource penetration testing framework used by several cybersecurity experts, including white-hat and black-hat 5 https:// apps. micro soft. com/ store/ apps/ windo ws.6 www. metas ploit. com/. Page 9 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 hackers, that allows executing security testsand developing and testing payloads and exploit codes. MF contains multiple encoders, Shikata Ga Nai and XOR encoders being the most representatives. Also, MF provides different tools such as msfvenom, an open-source payload generator that combines other tools like msfpayload and msfencode to create an obfuscated binary. Data processing DEFENDIFY relies on computer vision DL algorithms;thus, the primary processing that data require is their transformation into images, which can be consumed by convolutional models employing TL. This component can be implemented through different programming languages, following a procedure similar to the next one: (1) the sample is loaded and read byte by byte, storing each byte-associated numerical value in a list; (2) a default image size is calculated (width, height) depending on the list length, i.e., the file size; (3) a grayscale image is created by copying the bytes one by one into an empty image, with every byte corresponding to one image pixel. For example, suppose that we have a 4-byte binary file, and after reading its bytes, the following values are obtained: [0,64,128,255]. If we want to arrange these values into a 2×2 size image, we will obtain an integer matrix as the one shown in Eq.(1), which represents an image that is saved and used later to feed the models. The size of the image associated with each sample is determined using a custom function that uses the number of bytes in the sample as a parameter, as shown in Table2. Based on the received list of bytes, the function defines a corresponding image width, and its height is calculated by dividing the file size by the image width as proposed inBensaoud etal. (2020). (1) Img =  0 64 128 255  A Python library that can generate an image from a Python list of bytes is Pillow.7 Using this library, a binary image was created for each of the files included in the training dataset, as explained in “Dataset creation” section, obtaining three new training datasets: i) an image dataset obtained from samples obfuscated with Shikata Ga Nai, ii) an image dataset obtained from samples obfuscated with XOR, and iii) an image dataset obtained from samples not obfuscated. When DEFENDIFY gets deployed, this procedure would be applied to any sample that needs to be classified after passing the entropy tester described in “Entropy tester” section. For example, the gray-scale image in Fig.2 was generated using the procedure previously described over a randomly selected encoded binary file and is an example of the possible outputs that can be obtained with the algorithm. Model generation This module is responsible for training and validating a model to detect non-obfuscated and obfuscated malware. We propose different components to process and classify new software samples. These components are as follows and described next: Entropy Tester, Detection Model Selection and Training, and Model Validation. Table 2 Image width according to the sample size Binary file size (KiB) Image width (px) Less than 10 32 10–30 64 30–60 128 60–100 256 100–200 384 200–500 512 500–1000 768 More than 1000 1024 Fig. 2 Example of gray‑scale image obtained from an encoded sample 7 https:// pythonpillow. org/. Page 16 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 ( 31, 449 ∗ 2 ∗ 20% = 12, 579 ). The models were unable to process 54 of the 12,579 samples originally included in the testing dataset, resulting in a total number of classified samples of 12,525. In addition, classification results obtained for each model are summarized in Table10, which indicates the relation in terms of percentage for each predicted class, dividing the number of samples of each “actual class” by the number of samples of each “predicted class”, per each tested model. The results for each metric defined in “Model validation” section, calculated for each of the models considered are presented in Table11. As seen in the testing results, all the models present noticeable good results. However, ResNet18 stands out Fig. 7 Confusion matrices for XOR and Shikata Ga Nai samples Table 10 Results obtained in the classification of XOR and Shikata Ga Nai samples Predicted class Actual class. Goodware Malware Goodware ResNet18 95.36% ResNet18 4.64% ResNet34 95.34% ResNet34 4.66% EfficientNetB3 95.44% EfficientNetB3 4.56% EfficientNetV2S 94.93% EfficientNetV2S 5.07% Malware ResNet18 0.57% ResNet18 99.43% ResNet34 0.60% ResNet34 99.40% EfficientNetB3 0.66% EfficientNetB3 99.34% EfficientNetV2S 0.57% EfficientNetV2S 99.43% Table 11 Metrics obtained in the classification of XOR and Shikata Ga Nai samples Bold values are the highest values obtained per each metric Model Metric Accuracy (%) Precision (%) Recall (%) F1 score (%) ResNet18 97.42 95.65 99.43 97.5 ResNet34 97.39 95.6 99.4 97.46 EffientNetB3 97.41 95.72 99.34 97.49 Efficient‑ NetV2S 97.21 95.25 99.43 97.3 Page 17 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 because it is the smallest network in terms of components, compared to other models considered, while maintaining a 97.42% of accuracy. For such a reason, ResNet18 could be considered as the best option to build a classifier that can differentiate malware from goodware with XOR and Shikata Ga Nai obfuscated binaries. Computational costs An analysis of computational costs for our proposed solution was done for each of the four CNN architectures presented before. The experiments consist of classifying a batch of images with a size varying between 1 samples and 104 Kb samples, while measuring the time and memory used to fulfill the task. To reduce randomness in the results, each batch size was tested with every model 20 times. In Figs.8 and 9, the measurements of time and RAM are presented, with the mean of all 20 repetitions highlighted as the darker lines, and the 95% confidence interval for these repetitions is drawn as the lines’ light contours. Figure8 shows the inference time as a linear function with the number of processed samples as the independent variable, and the slope depends on the CNN architecture used. In particular, we can observe that the smallest line slope is associated with ResNet18, which means that this model is the fastest. Besides, the linear behavior indicates stability, because feeding more samples to the program will not make it run indefinitely for many seconds. Figure9 shows memory consumption as a logarithmiclike function, although when the number of samples is large enough (more than 1000 approximately) it can be considered constant in practice, with a value depending on the CNN architecture. For example, ResNet18 has the lowest RAM usage with a value of approximately 0.47 GiB. As the growth of memory consumption is too slow to be relevant, the proposed solution is also stable in relation to RAM usage. Finally, it is worth noting that the 95% confidence interval is too small even to be clearly seen in the graphics. This means that the experiment results are trustworthy and can be easily reproducible in a similar computational environment. Prototyping DEFENDIFY Modern and visionary societies need to be built over strong foundations, and in this regard, NSAI (Network System of Artificial Intelligence) is one of the most prominent keystones(Song etal. 2022). NSAI is a paradigm where AI is immersive in all components of a network system, including edge devices, communication nodes and on-premise or cloud-based servers. NSAI architecture is composed of 4 tiers: physical network (PN), service-customized network (SCN), generalized smart service (GSS), and application (APP). In particular, GSS is one of the most critical tiers, as it may be seen as an intermediate tier that facilitates applications to operate without being aware of the complexity of communication and computing below. To achieve the previous goal, GSS provides application programming interfaces (API) to be consumed by applications in the upper tier. GSS also has the autonomy to reconfigure upper and lower components, i.e. APP and SCN settings, to guarantee reliable services in a smart city. DEFENDIDY may be integrated into GSS as a cybersecurity service as it provides an AI-enabled smart service capable of detecting obfuscated malware, independent of its location in a distributed network. This service may work by automating the detection of malware flowing to user devices, or it may also be consumed on demand through a smart service API that allows users to upload files to it and trigger a classification process as Fig. 8 Time consume, in seconds, registered by each CNN model Fig. 9 RAM consume, in GiB, registered by each CNN model Page 18 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 described in “DEFENDIFY framework” section. DEFENDIFY would inform users about the analysis results and the probabilities that the provided sample is goodware or malware. Also, it might be appropriate to have the storage available to save new malware samples uploaded by users, allowing the database to grow over time, improving the detection models. For the sake of contributing a functional prototype that shows in an interactive way the functioning of our proposal, the deployment of DEFENDIFY was tested with a minimalist private web application made using the Python dash11 library. This web application consists of an on-demand simple interface where the user can upload an executable file, that is then processed as indicated in the data flow depicted in Fig.1. Figure10 shows a screenshot of the web application12 that runs the proposal described in this paper. Furthermore, DEFENDIFY’s lightweight architecture and efficient models, such as ResNet18, make it wellsuited for potential deployment on resource-constrained edge devices commonly found in smart city environments. For example, the framework could be adapted to run on platforms like Raspberry Pi 5 or NVIDIA Jetson Orin, which are popular choices for IoT and edge AI applications. Deploying DEFENDIFY on these edge devices would allow real-time malware detection closer to the data source, reducing latency, and enhancing the overall security posture of smart city networks. Discussion This work presents DEFENDIFY, a novel framework for detecting obfuscated malware using computer vision and deep learning techniques. The results demonstrate that DEFENDIFY is able to detect both obfuscated and nonobfuscated malware samples accurately. Compared to prior works, DEFENDIFY is the first framework to incorporate an entropy tester to differentiate between obfuscated and non-obfuscated binaries. This allows the framework to selectively use one of two deep learning models depending on whether obfuscation is detected. Other works such asConti etal. (2022) attempt to detect obfuscation in Android applications using deep learning, but do not provide a complete framework solution. The DL architectures explored in DEFENDIFY leverage transfer learning to avoid expensive training procedures and enable high accuracy even with limited training data. The ResNet18 architecture achieved the best results, detecting non-obfuscated and obfuscated malware with F1 scores of 99.34% and 97.5%, respectively. This compares favorably with related work such as Ravi et al. (2022) and Conti etal. (2022), reporting malware detection accuracy between 95–98% A key advantage of DEFENDIFY is the ability to detect malware obfuscated with real-world encoders such as XOR and Shikata Ga Nai that are commonly used by attackers(Burita and Le 2021; Webroot 2020). Fig. 10 Web interface with the deployment of DEFENDIFY 11 https:// pypi. org/ proje ct/ dash/. 12 https:// huggi ngface. co/ spaces/ esab/ malwa re_ detec tion. Page 19 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 Most prior work does not evaluate on such realistically obfuscated data. Testing on these encoders makes the framework more robust to new obfuscation techniques employed by malware creators. Transfer Learning (TL) offers significant advantages over training deep learning models from scratch for obfuscated malware detection. Recent research using non-TL approaches, such as depth-wise CNN with attention mechanisms (Han et al. 2024) and dynamic analysis CNN (Mercaldo et al. 2023), have demonstrated the viability of traditional deep learning methods. However, our TL-based approach achieves superior results in terms of: • Reduced malware dataset: DEFENDIFY employs a relative small dataset of 31,449 images (goodware and malware), due to it is inheriting the data features of the ImageNet-1K dataset, which is composed of more than 1.2 millions of images. To train a CNN without TL requires millions of samples to achieve comparable results in terms of accuracy as the ones obtained by DEFENDIFY (99.34% for non-obfuscated samples and 97.42% for obfuscated samples). • Reduced training times: Using Transfer Learning may require a different training time depending on the number of epochs and the capabilities of the computational infrastructure. In the case of DEFENDIFY, thanks to TL it was possible to train a CNN in less than 5min with 4 epochs using a Nvidia Saturn Cloud T4-4XLarge instance with 64 GB RAM, 16 CPU, 1 GPU and 1 TiB disk space. This time metric could not be achieved if TL is not used, as the training could last weeks. • Less computational resources: Training a CNN with TL requires instances with significantly less computational features than training without TL, e.g. midrange GPUs vs multi-GPUs, Quad-core processors (Intel i5/i7, Ryzen 5/7) vs High-performance CPUs (Intel i9, Ryzen 9, Threadripper), 16 Gb vs 32 Gb of RAM, 512 Gb vs 1-2 Tb of storage. The effectiveness of TL can be attributed to leveraging pre-trained weights from large-scale image datasets, which provide robust feature extraction capabilities. This transfer of knowledge is particularly valuable in malware detection, where acquiring large-scale labeled datasets is challenging. Our results demonstrate that TL can effectively bridge these practical limitations while maintaining or improving detection performance compared to traditional approaches, offering a compelling balance of performance, efficiency, and practicality for obfuscated malware detection. The computational analysis demonstrates that DEFENDIFY can scale efficiently thanks to the lightweight ResNet18 model. The throughput exceeds 500 samples per second with under 500MB memory consumption on a GPU instance. This efficiency enables deployment to production environments for practical malware detection. At last we may also consider the application of DEFENDIFY in the context of a smart city, where a malware campaign is addressed to some critical smart city service, like a traffic management service, impacting the dynamics of a smart city, due it would decrease the productivity of business that depends on transport like logistics and cargo transport and would increase the pollution, just to mention some effects. Considering a more individual sphere, we could also evaluate the impact that would be generated by a malware campaign addressed to impact a mobile transport application used by smart citizens to consult the schedule of buses or trains or review best routes to reach a destination. As generally, the end user device is the one more vulnerable between the components of a system, probably the malware campaign would try to exploit a vulnerability associated with such device to affect directly the transport application, e.g. through an abuse of the device administration API (T1616)(Mitre 2025a), or to affect services consumed by the transport application, e.g. altering the domain dynamic resolution (T1637)(Mitre 2025b). In this case, DEFENDIFY may detain the implantation of the malware or some of their components through a threat analysis performed directly in the user device or in a centralized service. Limitations Our results show the effectiveness of using CNN and TL for malware detection, however, these also bring some limitations in our proposal. First, a limitation coming from the use of CNN architectures is that its hierarchical feature extraction may not perfectly align with the obfuscated malware’s structural patterns. Second, TL uses pre-trained weights from natural image datasets like ImageNet, which may not be optimal to capture the unique characteristics of binary-derived images. These two limitations may impact our proposal ability to detect certain obfuscation patterns. These limitations are addressed by the properties of each one of the CNN architectures tested in the DEFENDIFY design: ResNet and EfficientNet. On the one hand, ResNet architecture uses residual connections to mainly address the vanishing gradient problem; however, such connections also preserve the fine-grained patterns existing in binaries, such as the ones that DEFENDIFY analyzes. On the other hand, EfficientNet uses Compound Scaling mainly to scale a network architecture in Page 20 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 a balanced way in terms of depth, width, and resolution. In that way, Compound Scaling allows a balance between accuracy and computational cost, but it also contributes in the ability of the CNN to extract features at different spacial resolutions, i.e. multi-scale feature extraction, which is crucial to recognize unique malware characteristics and patterns independent of the imagebinary size. Despite these inherent limitations, our experiments depict a robust performance across obfuscated and nonobfuscated samples, ResNet obtains the highest Accuracy (99.34% for Non obfuscated samples and 97.42% for Obfuscated samples), and EfficientNet obtains the highest Precision (99.53% for Non obfuscated samples and 95.72% for Obfuscated samples). Both architectures successfully learn hierarchical binary pattern representations while maintaining efficiency through residual connections (ResNet) and compound scaling (EfficientNet). Another limitation is the framework’s performance against zero-day malware attacks, which by definition use novel techniques not seen in the training data. Thus, the current implementation may struggle to detect completely new obfuscation methods or malware families. Additionally, the framework’s adaptability to different computing environments, such as resource-constrained IoT devices or high-performance servers, requires further investigation. To address these limitations, as future work we plan to explore continual learning approaches to adapt the models to new threats over time. Techniques like few-shot learning could also be investigated to improve detection of novel malware families with limited samples. Regarding adaptability, developing lightweight versions of the framework optimized for different hardware targets would enhance its practical applicability across diverse computing environments. In summary, DEFENDIFY pushes the state-of-theart in obfuscated malware detection through a robust framework combining an entropy tester, deep transfer learning models, and evaluation on realistically encoded malware data. Key advantages include high detection accuracy, flexibility to different obfuscation techniques, and efficient computation for practical purposes. While challenges remain, particularly for zero-day attacks and diverse computing environments, DEFENDIFY provides a strong foundation for future research in adaptive and robust malware detection systems. Conclusions andfuture work Malware continues to pose a critical threat to the security of computers, mobile phones and IoT devices(Díaz-López etal. 2018). These malicious programs can spy on users, steal sensitive data and even encrypt files on infected devices(SánchezVenegas etal. 2019). Moreover, in response to advanced protective measures, malware increasingly employs obfuscation techniques that hinder detection, making it challenging to distinguish malicious software from benign programs. In this context, traditional Machine Learning (ML) and Deep Learning (DL) techniques have proven ineffective and the research community must shift focus towards new methods capable of detecting obfuscated malware. In this paper, we introduced DEFENDIFY, a framework that demonstrates the potential of using computer vision and deep learning techniques for obfuscated malware detection. By converting malware binaries into images and applying transfer learning, DEFENDIFY achieves high accuracy in detecting obfuscated malware samples. This indicates that visual malware analysis combined with deep learning constitutes a promising approach for overcoming the limitations of conventional malware detection methods in handling obfuscated malware. DEFENDIFY consists of three main components, namely: Dataset Creation, Binary Obfuscation, and Model Generation. Specifically, the Dataset Creation module collects both malware and goodware samples. The the Binary Obfuscation module then selects specfic encoding mehods to obfuscate the malware. After the encoders are chosen, the module obfuscates the samples, transforms them into greyscale images and creates three training sets. The first one contains Shikata Ga Nai obfuscated malware and goodware, the second contains XOR obfuscated samples, and the third includes nonobfuscated samples. Finally, the Model Generation module incorporates an entropy tester to determine whether samples are obfuscated. Leveraging this framework, two DL models powered with transfer learning and previously trained, are employed. The first model detects malware from non-obfuscated samples, while the second model is used for obfuscated samples. To effectively apply the TL strategy, DEFENDIFY utilizes pre-trained weights for the feature extractors layers and trains only the last classification layers. Finally, performance is evaluated in DEFENDIFY using Accuracy, Precision, Recall, and F1-score metrics. Additionally, we validated the performance and computational efficiency of our approach using real-world malware and goodware samples. To this end, the Model Generation module was configured to test four DL architectures: ResNet18, ResNet 34, EfficientNetB3, and EfficientNetV2S. In terms of evaluation metrics, we found that ResNet18 outperformed the other architectures for both obfuscated and non-obfuscated samples, achieving F1-scores of 99.34% and 97.5%, respectively. Regarding computational consumption, RestNet18 required the least resources, followed by RestNet34, EfficientNetB3, and finally EfficientNetV2S. Page 21 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 As future work, we plan to explore interpretability techniques to identify the specific location of malicious code within the malware samples. These techniques will enhance trust in the model by not only classifying samples as malware or goodware, but also providing explanations for why a sample is categorized in a particular way. Additionally, we aim to address certain limitations of the framework by improving the diversity of the datasets used, ensuring broader coverage of different obfuscation techniques and malware families. This will enhance the model’s adaptability to emerging threats. Furthermore, we plan to investigate methods to strengthen the framework’s effectiveness against ever-evolving obfuscation techniques, ensuring its applicability to diverse malware types. An additional limitation of our current work is that it focuses solely on x86 binaries. Exploring crossarchitecture compatibility, such as assessing ARM binaries on x86 platforms, is an important direction for future research. We also plan as future works to develop experiments to perform a comparison between the method (Binaryto-Pixel Mapping) to convert binary files into images employed currently by DEFENDIFY, with other methods like: conversion of instructions (opcodes) to images, visualization based on entropy, generation of colored images using a RGB mapping, conversion of individual Portable Executable (PE) sections, among others. Besides, it could prove useful to implement algorithms that do not rely on binary to image transformation processes, using traditional neural network architectures that can handle unprocessed binary information, in order to compare the results obtained when images are not used. In this way, processing pipelines similar to DEFENDIFY, but that do not use TL or CNNs, may be developed as a future work. Acknowledgements We acknowledge all colleagues from School of Engineering, Science and Technology at Universidad del Rosario (Colombia) and Faculty of Computer Science at Universidad de Murcia (Spain) who provided feedback to improve the quality of this paper. Author contributions Conceptualization: Rodrigo Castillo Camargo, Santiago Alférez, Daniel Díaz‑ López; Methodology and validation: Angel Luis Perales Gómez, Pantaleone Nespoli; Data curation, formal analysis and investigation: Juan Murcia Nieto, Rodrigo Castillo Camargo, Nicolás Rojas; Writing ‑ original draft prepara‑ tion: Juan Murcia Nieto, Rodrigo Castillo Camargo, Nicolás Rojas, Pantaleone Nespoli; Writing ‑ review and editing: Daniel Díaz‑López, Félix Gómez Mármol, Umit Karabiyik; Funding acquisition: Daniel Díaz‑López, Angel Luis Perales Gómez, Pantaleone Nespoli, Santiago Alférez; Supervision: Santiago Alférez, Daniel Díaz‑López. Funding This work has been partially funded by the School of Engineering, Sci‑ ence and Technology at Universidad del Rosario (Colombia) through a “Beca de Estancia de Docencia e Investigación ‑ EDI 2022‑1”, by MCIN/AEI/ 10.13039/501100011033, NextGenerationEU/PRTR, UE, under Grant TED 2021‑ 129300B‑I00, by MCIN / AEI / 10.13039 / 501100011033 / FEDER, UE, under Grant PID2021 ‑ 122466OB ‑ I00, the strategic project DEFENDER from the Spanish National Institute of Cybersecurity (INCIBE), by the Recovery, Transfor‑ mation and Resilience Plan, Next Generation EU, and by the Spanish Ministry of Universities linked to the European Union through the NextGenerationEU program, under Margarita Salas postdoctoral fellowship (172/MSJD/22). This work has also been partially funded by the Nvidia Academic Grant Program through GPU instances provided by Saturn Cloud. Availability of data and materials Instructions on how to obtain datasets used in the experiments are available at the project repository at https:// github. com/ Nirogu/ Obfus cated Malwa reDet ection. Code availability Code developed to execute the experiments presented in this paper is avail‑ able at the project repository at https:// github. com/ Nirogu/ Obfus cated Malwa reDet ection. Declarations Competing interests The authors declare no competing interests. Received: 16 May 2024 Accepted: 17 March 2025 References Allix K, Bissyandé TF, Klein J et al (2015) Are your training datasets yet relevant? In: Piessens F, Caballero J, Bielova N (eds) Engineering secure software and systems. Springer, Cham, pp 51–67 Balram N, Hsieh G, McFall C (2019) Static malware analysis using machine learning algorithms on APT1 dataset with string and PE header features. In: 2019 International conference on computational science and compu‑ tational intelligence (CSCI). IEEE, pp 90–95 Bat‑Erdene M, Park H, Li H et al (2017) Entropy analysis to classify unknown packing algorithms for malware detection. Int J Inf Secur 16(3):227–248. https:// doi. org/ 10. 1007/ s10207‑ 016‑ 0330‑4 Ben‑Bassat M, Klove KL, Weil MH (1980) Sensitivity analysis in Bayesian clas‑ sification models: multiplicative deviations. IEEE Trans Pattern Anal Mach Intell PAMI 2(3):261–266. https:// doi. org/ 10. 1109/ TPAMI. 1980. 47670 15 Bensaoud A, Abudawaood N, Kalita J (2020) Classifying malware images with convolutional neural network models. Int J Netw Secur 22(6):1022–1031 Bergstra J, Bardenet R, Bengio Y, et al (2011) Algorithms for hyper‑parameter optimization. In: Proceedings of the 24th international conference on neural information processing systems. Curran Associates Inc., Granada, Spain, NIPS’11, pp 2546–2554 Breiman L (2001) Random forests. Mach Learn 45:5–32. https:// doi. org/ 10. 1023/A: 10109 50718 922 Burita L, Le DT (2021) Cyber security and apt groups. In: 2021 Communication and information technologies (KIT), pp 1–7. https:// doi. org/ 10. 1109/ KIT52 904. 2021. 95837 44 Ceschin F, Botacin M, Lüders G, et al (2021) No need to teach new tricks to old malware: winning an evasion challenge with xor‑based adversarial sam‑ ples. In: Reversing and offensive‑oriented trends symposium. Association for Computing Machinery, Vienna, ROOTS’20, pp 13–22. https:// doi. org/ 10. 1145/ 34336 67. 34336 69 Conti M, Vinod P, Vitella A (2022) Obfuscation detection in android applications using deep learning. J Inf Secur Appl 70:103311 Cover T, Hart P (1967) Nearest neighbor pattern classification. IEEE Trans Inf Theory 13(1):21–27. https:// doi. org/ 10. 1109/ TIT. 1967. 10539 64 Cox D (1989) Analysis of binary data, 2nd edn. Monographs on statistics and applied probability. Chapman and Hall, London Cyberfish (2025) Zero phishing protection for MSPs and SMB | cofense protect. https:// cofen se. com/. Accessed 15 Mar 2025 Díaz‑López D, Blanco Uribe M, Santiago Cely C et al (2018) Developing secure IoT services: a security‑oriented review of IoT platforms. Symmetry. https:// doi. org/ 10. 3390/ sym10 120669 Page 22 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 El Merabet H, Hajraoui A (2019) A survey of malware detection techniques based on machine learning. Int J Adv Comput Sci Appl 10(1):366–373 Farley R, Wang X (2014) Codext: automatic extraction of obfuscated attack code from memory dump. In: Chow SSM, Camenisch J, Hui LCK et al (eds) Inf Secur. Springer, Cham, pp 502–514 Fisher RA (1936) The use of multiple measurements in taxonomic problems. Ann Eugen 7(2):179–188. https:// doi. org/ 10. 1111/j. 1469‑ 1809. 1936. tb021 37.x Géron A (2019) Hands‑on machine learning with scikit‑learn, keras, and tensorflow. O’Reilly Media Inc, Newton Glaroudis D, Iossifides A, Chatzimisios P (2020) Survey, comparison and research challenges of IoT application protocols for smart farming. Comput Netw 168:107037 Graham D (2021) Ethical hacking: a hands‑on introduction to breaking in. No Starch Press, San Francisco Guo Y, Fan W (2019) Feature collection and selection in malware classifica‑ tion. In: Proceedings of the 2019 international conference on artificial intelligence and advanced manufacturing, pp 1–5 Hakon P, Fareed Y, Sindre G (2020) The ransomware‑as‑a‑service economy within the darknet. Comput Secur 92:101762. https:// doi. org/ 10. 1016/j. cose. 2020. 101762 Han S, Yun H, Park Y (2024) Deep learning for cybersecurity classification: utilizing depth‑wise CNN and attention mechanism on VM‑obfuscated data. Electronics 13(17):3393 Hussain A, Asif M, Ahmad MB, et al (2022) Malware detection using machine learning algorithms for windows platform. In: Proceedings of interna‑ tional conference on information technology and applications: ICITA 2021. Springer, pp 619–632 Iaksch J, Fernandes E, Borsato M (2021) Digitalization and big data in smart farming—a review. J Manag Anal 8(2):333–349 Jiang Y, Li R, Tang J et al (2020) Aomdroid: detecting obfuscation variants of android malware using transfer learning. In: Park N, Sun K, Foresti S et al (eds) Security and privacy in communication networks. Springer, Cham, pp 242–253 Joshi S, Upadhyay H, Lagos L, et al (2018) Machine learning approach for malware detection using random forest classifier on process list data structure. In: Proceedings of the 2nd international conference on information system and data mining, pp 98–102 Kakisim AG, Gulmez S, Sogukpinar I (2022) Sequential opcode embedding‑ based malware detection method. Comput Electr Eng 98:107703 Ke G, Meng Q, Finley T, et al (2017) Lightgbm: a highly efficient gradient boosting decision tree. In: Advances in neural information processing systems, vol 30. Curran Associates, Inc., Long Beach Kim JY, Bu SJ, Cho SB (2018) Zero‑day malware detection using transferred generative adversarial networks based on deep autoencoders. Inf Sci 460–461:83–102. https:// doi. org/ 10. 1016/j. ins. 2018. 04. 092 Lee H, Kim S, Baek D et al (2023) Robust IoT malware detection and clas‑ sification using opcode category features on machine learning. IEEE Access 11:18855–18867 Li M, Han D, Li D et al (2022) MFVT: an anomaly traffic detection method merging feature fusion network and vision transformer architecture. EURASIP J Wirel Commun Netw 1:39 Lyda R, Hamrock J (2007) Using entropy analysis to find encrypted and packed malware. IEEE Secur Priv 5(2):40–45. https:// doi. org/ 10. 1109/ MSP. 2007. 48 Ma C (2021) Smart city and cyber‑security; technologies used, leading chal‑ lenges and future recommendations. Energy Rep 7:7999–8012 Marastoni N, Giacobazzi R, Dalla Preda M (2021) Data augmentation and transfer learning to classify malware images in a deep learning context. J Comput Virol Hack Techn 17(4):279–297. https:// doi. org/ 10. 1007/ s11416‑ 021‑ 00381‑3 Martínez Martínez I, Florián Quitián A, Díaz‑López D et al (2021) Malseirs: forecasting malware spread based on compartmental models in epidemiology. Complexity 2021:5415724. https:// doi. org/ 10. 1155/ 2021/ 54157 24 Mercaldo F, Ciaramella G, Santone A, et al (2023) Obfuscated mobile malware detection by means of dynamic analysis and explainable deep learning. In: Proceedings of the 18th international conference on availability, reliability and security, pp 1–10 Mimecast (2025) Cybergraph datasheet–mimecast. https:// www. mimec ast. com/ resou rces/ datas heets/ cyber graph/. Accessed 15 Mar 2025 Mitre (2025a) Abuse elevation control mechanism: device administrator permissions. https:// attack. mitre. org/ techn iques/ T1626/ 001/. Accessed 15 Mar 2025 Mitre (2025b) Dynamic resolution. https:// attack. mitre. org/ techn iques/ T1637/. Accessed 15 Mar 2025 Nespoli P, Useche Peláez D, Díaz‑López D et al (2019) Cosmos: collabora‑ tive, seamless and adaptive sentinel for the internet of things. Sensors. https:// doi. org/ 10. 3390/ s1907 1492 Pelaez DU, Díaz‑López D, Nespoli P, et al (2018) Tris: a three‑rings IoT sentinel to protect against cyber‑threats. In: 2018 Fifth international conference on internet of things: systems, management and security, pp 123–130. https:// doi. org/ 10. 1109/ IoTSMS. 2018. 85544 32 Pereira‑Kohatsu J, Quijano‑Sánchez L, Liberatore F et al (2019) Detecting and monitoring hate speech in twitter. Sensors 19(21):4654 Platt J (2000) Probabilistic outputs for support vector machines and com‑ parisons to regularized likelihood methods. Adv Large Margin Classif 10:61–74 Ravi V, Alazab M, Selvaganapathy S et al (2022) A multi‑view attention‑ based deep learning framework for malware detection in smart healthcare systems. Comput Commun 195:73–81 Rehman A, Saba T, Khan MZ et al (2022) Internet‑of‑things‑based suspicious activity recognition using multimodalities of computer vision for smart city security. Secur Commun Netw 2022:8383461 Rezaei T, Manavi F, Hamzeh A (2021) A PE header‑based method for mal‑ ware detection using clustering and deep embedding techniques. J Inf Secur Appl 60:102876 Ribani R, Marengoni M (2019) A survey of transfer learning for convolutional neural networks. In: 2019 32nd SIBGRAPI conference on graphics, pat‑ terns and images tutorials (SIBGRAPI‑T). IEEE, pp 47–57 Ristvej J, Lacinák M, Ondrejka R (2020) On smart city and safe city concepts. Mob Netw Appl 25:836–845 Sahin M, Bahtiyar S (2020) A survey on malware detection with deep learn‑ ing. In: 13th international conference on security of information and networks, pp 1–6 Sánchez Venegas C, Aguado Bedoya C, Díaz‑López D et al (2019) Using reverse engineering to handle malware. Ing Solidar 15(2):1–26. https:// doi. org/ 10. 16925/ 2357‑ 6014. 2019. 02. 02 Shaukat K, Luo S, Varadharajan V (2023) A novel deep learning‑based approach for malware detection. Eng Appl Artif Intell 122:106030 Smith LN (2017) Cyclical learning rates for training neural networks. arXiv: 1506. 01186 Song L, Hu X, Zhang G et al (2022) Networking systems of AI: on the con‑ vergence of computing and communications. IEEE Internet Things J 9(20):20352–20381. https:// doi. org/ 10. 1109/ JIOT. 2022. 31722 70 Talukder S, Talukder Z (2020) A survey on malware detection and analysis tools. Int J Netw Secur Appl 12(2): 21 Tan M, Le Q (2019) EfficientNet: rethinking model scaling for convolutional neural networks. In: Chaudhuri K, Salakhutdinov R (eds) Proceedings of the 36th international conference on machine learning, proceedings of machine learning research, vol 97. PMLR, Long Beach, California, USA, pp 6105–6114. https:// proce edings. mlr. press/ v97/ tan19a. html Tan M, Le QV (2020) Efficientnet: rethinking model scaling for convolutional neural networks. arXiv: 1905. 11946 Tan M, Le QV (2021) Efficientnetv2: smaller models and faster training. arXiv: 2104. 00298 Toffalini F, Ochoa M, Sun J, et al (2019) Careful‑packing: a practical and scalable anti‑tampering software protection enforced by trusted computing. In: Proceedings of the ninth ACM conference on data and application security and privacy. Association for Computing Machinery, Richardson, Texas, USA, CODASPY ’19, pp 231–242. https:// doi. org/ 10. 1145/ 32920 06. 33000 29 Useche‑Pelaez DE, Sepulveda‑Alzate D, Díaz‑López D et al (2018) Building malware classificators usable by state security agencies. Iteckne 15:107–121 Webroot (2020) Webroot threat report. Tech. rep., Webroot. https:// mypage. webro ot. com/ rs/ 557‑ FSI‑ 195/ images/ 2020% 20Web root% 20Thr eat% 20Rep ort_ US_ FINAL. pdf Weiss K, Khoshgoftaar TM, Wang D (2016) A survey of transfer learning. J. Big Data 3(1):1–40 Page 23 of 23 CastilloCamargoetal. Cybersecurity (2025) 8:97 Xi B (2020) Adversarial machine learning for cybersecurity and computer vision: current developments and challenges. Wiley Interdiscip Rev: Comput Stat 12(5):e1511 Xue H, Sun S, Venkataramani G et al (2019) Machine learning‑based analysis of program binaries: a comprehensive study. IEEE Access 7:65889–65912 Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations.