Does Reasoning Mitigate Backdoor Attacks? A Neuro-Symbolic Perspective

Neuro-Symbolic AI systems are not automatically secure despite their design goals: While these models were assumed to be "robust-by-design" due to their integration of reasoning with neural networks, the paper provides the first systematic evidence that they can still be vulnerable to backdoor attacks, challenging assumptions about their trustworthiness.

Reasoning processes can provide both protection and vulnerability: The study found that NeSy models show more robustness than standard neural networks on average, but this protection is inconsistent—it "vastly depends on the strictness of the reasoning process being enforced," meaning weak reasoning constraints may offer little defense against attacks.

Compatibility between attack goals and reasoning logic matters significantly: The robustness of NeSy systems varies based on "compatibility with the chosen adversarial target," suggesting that attackers who understand a system's reasoning rules may be able to craft more effective backdoor attacks that align with those logical constraints.

Additional complexity in AI systems can create new security risks: The paper identifies that the neural-symbolic integration process, while beneficial for transparency and explainability, "constitutes an additional layer of complexity that may provide an attack entry-point"—a reminder that adding components to improve AI safety requires careful security evaluation.

Systematic security testing of emerging AI paradigms is essential: This research demonstrates the importance of proactively evaluating new AI approaches for adversarial vulnerabilities rather than assuming theoretical benefits translate to practical security.