The BioQOP project deals with an architecture where a facial recognition system is distributed on several nodes.
The innovation is to train each of these nodes in such a way that:
(1) Individually, each node has a weak distinguishability capability;
(2) Taken as a whole, the performance of the system, once all results are gathered, is optimal in terms of biometric recognition accuracy (under the privacy constraints).
Biometric data processing is realized with deep convolutional neural networks (CNNs). Thanks to property (1), an adversary who has access to one node is not able to effectively exploit the underlying CNN signature for classifying biometric data. We plan to measure privacy loss thanks to Differential Privacy techniques. Furthermore, during the BioQOP project, we want to design optimization techniques for accessing and querying costly resources – the cost here being the required differential privacy budget.