The problem: vulnerability to intentional manipulation

As algorithms are integrated into critical infrastructure, the question is no longer simply whether they are accurate on sterile datasets, but whether they can maintain that accuracy under attack. In the realm of cybersecurity, Intrusion Detection Systems (IDS) face the threat of Adversarial Machine Learning (AML). Attackers analyze the behavior of the defensive algorithm and generate adversarial examples mathematically designed to evade detection. If a security algorithm implicitly assumes that input data comes from a benign distribution, it lacks the necessary resilience for the real world. A system that can be fooled by slight perturbations in its inputs fails in its fundamental purpose: it cannot justify institutional trust because it operates as an agent vulnerable to external co-option.

What do we propose in this work?

To address this challenge, we developed Apollon, a novel defense system to protect IDS against evasion attacks. Instead of relying on a single static model that an attacker can study and bypass, Apollon introduces strategic uncertainty into the defense.

The system utilizes a diverse set of classifiers and employs Multi-Armed Bandits (MAB) with Thompson sampling to dynamically select the optimal classifier or ensemble of classifiers for each network input. By evaluating Apollon on recent datasets, we demonstrate that this dynamic selection prevents attackers from learning the system’s behavior, neutralizing the generation of adversarial examples.

The connection to trustworthy algorithms

This work materializes the formal guarantees of the Technical Dimension of trustworthy algorithms, demonstrating that trust must be engineered into the system’s architecture:

• Robustness: Apollon maintains acceptable behavior and high detection performance even when input data is intentionally manipulated by an attacker. System degradation is prevented by breaking the assumptions of classical attack methods.
• Security: The algorithmic design actively resists attacks and misuse. By using Multi-Armed Bandits, the system prevents the attacker from inferring the underlying decision logic, protecting the IDS from evasion.

Read the full research

If you want to dive deeper into the mathematical architecture of Apollon, the implementation of Thompson sampling, or the results of the evaluations against white-box and black-box attacks, the article has been published in the journal Computers & Security.

👉 Read the full paper here (DOI: 10.1016/j.cose.2023.103546)