Depth-induced bipolar neural collapse for privacy-preserving face verification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural Collapse (NC) describes feature convergence in deep networks during supervised training.
• Existing research links NC to label-driven gradient optimization.
• A distinct, training-free regime called Bipolar Neural Collapse (BNC) is unexplored.

Purpose of the Study:

• To investigate a label-free, training-free mechanism for generating stable representations.
• To explore the potential of BNC for privacy-preserving face verification.
• To analyze the impact of transformation depth on BNC geometry and utility.

Main Methods:

• A fixed-weight multilayer transformation applied to pretrained facial embeddings.
• Utilized top-k selective routing and unit-norm projection without backpropagation.
• Evaluated on LFW, CFP, and CMU-PIE benchmarks with varying transformation depths.

Main Results:

• The transformation induces a low-rank bias, polarizing embeddings into antipodal clusters.
• This BNC geometry enables one-way cryptographic hashing for face verification without raw data storage.
• Moderate depth enhances verification utility, while excessive depth risks over-collapse and increased collisions.

Conclusions:

• Depth-induced BNC geometry offers a controllable, training-free approach to representation learning.
• It balances face verification utility with privacy-preserving protected-template security.
• Findings suggest BNC as a novel mechanism for secure and efficient biometric systems.