Unveiling Hidden Threats: Deep Learning Steganography for Malware Spread 🔍

Secrecy with Intent: Malware Propagation through Deep Learning-Driven Steganography Mikhail Diyachkov1, Arkadi Yakubov1, Hadassa Daltrophe1 and Kiril Danilchenko2, 1Shamoon College of Engineering, Israel, 2University of Waterloo, Canada Abstract With the proliferation of deep learning, steganography techniques can now leverage neural networks to imperceptibly hide secret information within digital media. This presents potential risks of propagating malware covertly. We present an innovative deep-learning framework that embeds malware within images for stealthy distribution. Our methodology transforms malware programs into image representations using a specialized neural network. These image representations are then embedded seamlessly within innocuous cover images using an encoding network. The resulting stego images appear unmodified to the naked eye. We develop a separate network to extract the malware from stego images. This attack pipeline allows the malware to bypass traditional signature-based detection. We experimentally demonstrate the efficacy of our approach and discuss its implications. Our framework achieves high-fidelity reconstruction of embedded malware programs with minimal distortions in the cover images. We also analyze the impact of loss functions on concealment and extraction capacity. The proposed technique represents a significant advancement in AI-driven steganography. By highlighting an intriguing attack vector, our work motivates research into more robust defensive solutions. Our study promotes responsible disclosure by releasing the attack implementation as open-source. Keywords Intrusion Detection System, Controller Area Network, In-Vehicle Network, LSTM Full Text : https://aircconline.com/csit/papers/vol14/csit141102.pdf Abstract URL : https://aircconline.com/csit/abstract/v14n11/csit141102.html Volume URL : https://airccse.org/csit/V14N11.html #intrusiondetection #lstm #security #blockchain #cybersecurity #deeplearning