rCamInspector: Building Reliability and Trust on IoT (Spy) Camera Detection using XAI

XAI tools are essential for building trust in AI security systems: The paper demonstrates that traditional methods of understanding why AI models make decisions (like mutual information analysis) can be unreliable. By using explainability tools like SHAP and LIME, the rCamInspector system provides transparent explanations for its security decisions, which is critical when AI systems are used to protect networks and identify potentially dangerous devices.

High accuracy alone doesn't guarantee trustworthy AI: Although the system achieved 92-99% accuracy in detecting IoT cameras, the researchers found that this high performance didn't automatically translate to reliable explanations. This highlights an important principle: responsible AI requires both accurate predictions AND the ability to explain why those predictions were made.

Different features matter for different decisions: The study found that the same AI model uses different network traffic features to make different types of decisions (e.g., "Init Bwd Win Byts" was most important for identifying certain camera types). This finding suggests that explainability isn't one-size-fits-all—understanding AI decisions requires examining the specific context and type of prediction being made.

Practical security benefits from explainability: By making the camera detection system explainable, network administrators and security teams can better understand and verify whether the system is making sound decisions, rather than blindly trusting a "black box" that might flag legitimate devices or miss actual threats.