Reproduce “DeepTrust: Multi-Step Classification through Dissimilar Adversarial Representations for Robust Android Malware Detection” on OpenResearch

About this paper

Daniel Pulido-Cortázar, Daniel Gibert, Felip Manyà

DeepTrust is a multi-step classification metaheuristic designed for robust Android malware detection, which maximizes the representational divergence among its internal deep neural network models. This framework secured a gold medal in the IEEE SaTML’25 ELSA-RAMD competition by demonstrating superior resilience against both feature-space and problem-space adversarial attacks, while maintaining high detection rates and a low false positive rate.

Key insights

• Maximizing the divergence of learned representations among sequential classifiers makes the decision space unpredictable for attackers, which significantly enhances system robustness against evasion attacks.
• Adversarial training, particularly when perturbing 'topk' features with the largest loss gradients, substantially improves model robustness without compromising accuracy on clean data.
• The multi-step strategy with diverse representations consistently outperforms individual models and traditional ensemble methods under adversarial conditions by presenting a more complex and unpredictable defense surface to attackers.

Summary by alphaXiv.