Securing Robot Vision: Leveraging Adversarial Detection for Data Poisoning Defense

Newswise — Robot vision systems are indispensable for autonomous functionalities such as object recognition and navigation, yet they remain highly vulnerable to data poisoning and adversarial attacks. While adversarial defenses have matured significantly, research into data poisoning—especially within the context of robotics—has lagged. Most current poisoning defenses are designed for specific attack vectors and lack the generalizability required for dynamic environments. For robots relying on vision-based models, poisoning attacks can corrupt training data to cause catastrophic misidentifications or navigation failures. This lack of versatile and robust defense mechanisms against data manipulation represents a critical technical pain point that threatens the reliable deployment of autonomous systems in safety-critical sectors.

To resolve this, the Nanjing University team proposed an innovative framework that adapts adversarial example detection for data poisoning defense. They identified key similarities between poisoning and adversarial examples regarding feature space distribution and sensitivity to model mutations. The team enhanced the Feature Squeezing method by incorporating grayscale compression, data augmentation, and dimensionality reduction techniques like PCA to filter out malicious noise. Additionally, they refined Model Mutation techniques to detect anomalies based on samples’ sensitivity variance. Experimental evaluations across various attack types, including Label Flipping and Targeted Clean-Label attacks, demonstrate that this cross-domain approach significantly outperforms traditional methods like De-Pois and Sever. This research provides a robust security layer for future autonomous robotics, bridging the gap between adversarial and poisoning protection.

https://link.springer.com/article/10.1007/s11704-025-50195-5