Explanation Residuals as Adversarial Tripwires for Encrypted-Traffic IDS

Extend Nguyen et al.’s MAIP framework (ARES 2023) from using XAI for post-hoc interpretability to making explanation-space deviations a first-class defense signal. Concretely, learn a deep anomaly detector with an auxiliary “explanation consistency” head that enforces expected attribution patterns (over features or latent channels) for normal traffic. At inference, compute an explanation residual (distance between a sample’s attributions and the expected attribution prototype for its class/subtype); large residuals flag adversarial or anomalous behavior even when the label looks normal. Most current IDS defenses monitor input or latent features; few treat explanation drift as the primary anomaly channel. This idea adapts explanation-based discrepancy to encrypted traffic flows, where raw content is hidden and features encode higher-level behavior. Addresses Sorensen et al. (SATC 2025) showing one-class models in IoT are brittle to FGSM/PGD by providing an orthogonal, model-agnostic detection layer that reacts to adversarial “why” changes, not just “what” changes. Aligns with surveys calling for robust, explainable defenses with real-time viability. Adversarial examples often preserve the predicted label while subtly altering gradient salience; this “explanation gap” is a rich deviation signal in encrypted settings. The approach is compatible with black-box models and ensembles. Impact: A practical path to trustworthy XAI in IDS—turning interpretability from a reporting tool into a detection mechanism that helps catch stealthy evasion on encrypted traffic.

References:

1. A deep learning anomaly detection framework with explainability and robustness. Manh-Dung Nguyen, Anis Bouaziz, Valeria Valdés, Ana Rosa Cavalli, Wissam Mallouli, Edgardo Montes de Oca (2023). ARES.
2. Adversarial Evasion Attacks on OCC-Based Machine Learning Intrusion Detection Systems in the Internet of Things. David Lykke Sorensen, Mohamed Baza, Mahmoud M. Badr, Tara Salman, Amar Rasheed (2025). 2025 1st International Conference on Secure IoT, Assured and Trusted Computing (SATC).