Privacy-Aware Traffic Re-Identification with Interpretable Sparse Autoencoders

Joris Heemskerk, Martijn Folmer, Michael Dubbeldam, Rianne van Os, Henry Maathuis

Track: Type A (Regular Papers)

Keywords: Mechanistic Interpretability, Sparse Autoencoder, Open-World Re-Identification

Abstract: Artificial Intelligence (AI) systems have become increasingly powerful tools used by businesses and governments to increase productivity, earnings, efficiency, and more. With this significant power comes the responsibility of balancing innovation with ethical considerations. We aim to increase transparency and trust in Computer Vision embedding models by suppressing unwanted identifiable features from a model trained on re-identification of traffic participants. This is achieved by using Sparse Autoencoders as a dictionary learning technique to extract highly interpretable features from our model. Unwanted identifiable features are suppressed and we analyse the effects on performance. Using this technique, we demonstrate that it is possible to create a transparent and highly interpretable model with a limited reduction in performance (a decrease from 0.98 mAP@0.6 to 0.90 mAP@0.6).

Serve As Reviewer: ~Henry_Maathuis1

Submission Number: 12