Illuminating the Black Box: Real-Time Monitoring of Backdoor Unlearning in CNNs via Explainable AI

Backdoor attacks are a serious security threat to AI systems: The paper demonstrates that without intervention, neural networks can be compromised to misclassify inputs when exposed to hidden "trigger" patterns, with attack success rates reaching 96.51% in their experiments.

Current unlearning methods lack transparency: Existing techniques to remove backdoor behaviors from AI models operate as "black boxes," making it difficult for organizations to verify that malicious triggers have actually been removed—a critical gap for responsible AI deployment.

Real-time monitoring of AI "unlearning" is now possible: By integrating explainable AI techniques (specifically Grad-CAM visualization), the researchers developed a framework that allows stakeholders to watch and verify in real-time how a compromised model is being cleaned, shifting from hidden processes to observable ones.

Effective backdoor removal doesn't require sacrificing overall performance: The framework successfully reduced attack success rates by 94.28% while maintaining 99.48% of the model's legitimate accuracy (82.06%), demonstrating that security improvements don't necessitate major performance trade-offs.

Quantifiable metrics enable verification of AI safety interventions: The introduction of the Trigger Attention Ratio (TAR) metric provides a measurable way to confirm that models are shifting their focus away from malicious triggers toward legitimate features, supporting verifiable and auditable AI safety practices.