Collaborative multi-instance feature aggregation for visual privacy protection

The rapid advancement of Deep Neural Networks (DNNs) has dramatically shaped numerous domains of modern society. With the increasing emphasis on high-quality and personalized data, safeguarding visual privacy and preventing information leakage have become urgent research priorities. Traditional privacy-preserving strategies predominantly rely on isolated instance-level manipulations, such as artificial noise injection or pixel-level obfuscation, to achieve de-identification. Yet, these independent sample-level approaches inevitably face an inherent trade-off: slight alterations provide inadequate protection, whereas aggressive modifications severely distort task-relevant semantics and compromise downstream accuracy. To overcome this dilemma, we introduce a novel visual privacy protection framework driven by multi-instance feature-level aggregation. Unlike conventional methods that perturb a single isolated instance, our approach constructs anonymized representations by collaboratively aggregating identity-discriminative feature fragments drawn from multiple data instances. Instead of relying on external noise to conceal identity, this collaborative aggregation synthesizes protected data, effectively weakening the deterministic link between the original identity and the visual representation. Comprehensive experiments on diverse visual identity recognition benchmarks demonstrate that the proposed framework delivers superior results, achieving both robust perceptual anonymization and high downstream task utility compared with prior state-of-the-art techniques. © 2026