Data Poisoning Attack Detection Based On Reconstruction

The application of artificial intelligence (AI) technologies in intelligent systems has become increasingly widespread. However, training high-performance AI models still relies on large-scale datasets, exposing them to data poisoning attacks. In this paper, we propose a reconstructionbased data poisoning detection technique to protect AI models from malicious manipulation, ensuring their security, robustness, and normal functionality. The method utilizes diffusion models to reconstruct input images and generate their “clean” counterparts. By incorporating an attention mechanism, the model focuses on critical features and patterns in the data. Due to the anomalous characteristics of poisoned samples, their attention weight distributions exhibit significant deviations. By comparing the differences between original and reconstructed data, the attention mechanism effectively identifies abnormal or maliciously altered samples. This approach integrates the reconstruction capability of diffusion models with the feature sensitivity of attention mechanisms, enabling reliable detection of poisoned samples and preserving the integrity of machine learning models during training. Experimental results demonstrate the method’s high detection accuracy and robustness across various attack scenarios.