A Deep Learning Approach for Binary Code Similarity Detection to Detect Vulnerabilities in Firmware Binaries

Connected devices are increasingly widespread and provide substantial advantages, yet they also face significant security challenges, making them susceptible to malware, denial-of-service (DoS), and network attacks, which can result in various issues such as data breaches, system malfunctions, and large-scale disruptions.Vulnerability detection in the connected devices is essential to safeguard against the myriad of security threats posed by the increasing number and diversity of connected devices.The implementation of robust vulnerability detection strategies helps to protect sensitive data, ensure the integrity of critical infrastructure, and maintain consumer trust.In this work, a novel vulnerability detection framework is proposed using graph attention network (GAT) and decodingenhanced bidirectional encoder representations from transformers with disentangled attention (DeBERTa) model.GAT is used for analyzing complex device networks and detecting vulnerabilities arising from device interactions.DeBERTa's advanced contextual understanding and feature extraction techniques help to detect complex vulnerabilities in connected devices, and its ability to handle diverse data makes it ideal for creating high-quality embeddings.Hence transformer-based language model and GAT are used to generate device binary embeddings.Binary embeddings are used in a generative adversarial networks model to classify it as vulnerable or normal.The dataset considered for the experiment, consists of 18 real-world vulnerabilities across 9 IoT firmware packages.The research findings show that proposed approach achieves good accuracy (by an average of 97.3%) based on real-world vulnerabilities across different firmware packages.Comparative analysis carried out shows the effectiveness of the proposed approach on existing vulnerability detection solutions, IoTSim, DeepWukong and Robin by substantially reducing the frequency of false alarms.