BFN (Backdoored Face-Networks Dataset)

This database is a database of backdoored neural networks intended for face recognition. The networks are of the FaceNet architecture and are trained on Casia-WebFace, with and without additional samples (which are the source of the backdoor). More information regarding backdoors and the project within which this fits can be found in the public release of the source code : https://gitlab.idiap.ch/bob/bob.paper.backdoored_facenets.biosig2022.

There are two sets of backdoored networks. A first one with backdoors with varying triggers (the triggers dataset) and a second one with backdoors with varying trigger placement strategies (the locations dataset). A third set of networks is also provided, just regular networks without any backdoor, referred to as the clean dataset. Configuration yaml files are provided to replicate the backdoored networks using the repository content linked above, in addition to pickle files containing validation scores on all validation datasets.

The purpose of this dataset is to allow for evaluation of backdoored network detection work on face recognition networks. The backdoored networks here are trained from a clean checkpoint and finetuned on poisoned data. The characteristics of each network is provided in the yaml file colocated with the checkpoint. Multiple triggers (organic and synthetic) are explored, in addition to multiple placement strategies (systematically random, static, in context etc). The finetuning is done exclusively on the layers reported in the yaml file.