New Handbook of Biometric Anti-Spoofing edition coming, researchers call for contributions
The 3rd Edition of the Handbook of Biometrics Anti-Spoofing (Presentation Attack Detection and Vulnerability Assessment) is planned for publication in late 2021 by Springer, which published the first and second editions, according to a Linkedin post by Marcel. Those editions were highly successful, with nearly 50,000 chapters downloaded, Marcel writes.
Marcel will be among the editors of the handbook, along with Julian Fierrez of the Universidad Autonoma de Madrid and Nicholas Evans of EURECOM.
The handbook will present the state-of-the-art in presentation attack detection for various biometric modalities, including face (visual, near infrared and thermal), voice, fingerprint, iris, vein, gait, handwriting or signature, and emerging modalities like behavioral and EEG biometrics.
New chapters will present new research in sensing and processing for PAD, as well as new techniques, comparative studies, surveys of the literature and standardization efforts, public datasets and new evaluation and assessment methodologies. Submissions about significant real-world PAD applications will also be considered.
Research contributions must be reproducible.
Authors are asked to express their interest by January 29, 2021, with a concise proposal. Chapters will be selected by February 26, 2021, with a final deadline for authors of June 30.
Vision transformers found effective for zero-shot PAD
Research into the effectiveness of vision transformers for ‘zero-shot face anti-spoofing’ by Marcel and Anjith George, also of Idiap, has also been published, suggesting a way to improve the effectiveness of detecting new spoof types.
Many PAD systems fail to generalize sufficiently to detect unseen attacks or work in environments they were not trained in, the researchers state in their paper ‘On the effectiveness of Vision Transformers for Zero-shot Face Anti-Spoofing.’ “Zero-shot” learning refers to machine learning tests against classes not used in training.
In testing against the HQ-WMCA and SiW-M datasets, the method outperformed state of the art techniques “by a large margin,” and also improved cross-database performance.
George and Marcel used transfer learning from a pre-trained Vision Transformer model. Face detection and landmark localization were performed using the MTCNN algorithm, the faces aligned and the images cropped during preprocessing, and a transformer model used to introduce self-attention layers, which “scan through and update each element in a sequence using the information from the whole sequence.” A Vision Transformer network described by Dosovitskiy et al. earlier this year was adapted, and the best model selected from among those trained based on the minimum loss from the validation set.
Models based on ResNetPAD, DenseNetPAD, and DeepPixBiS were used for comparison against the ViTranZFAS framework the researchers arrived at. In testing against the HQ-WMCA dataset with the BPCER set to 1 percent, the new framework had an ACER of 9.2, with a standard deviation of 7.99 percent, compared to an ACER of 15.55 with a standard deviation of 15.76 percent for DeepPixBiS, which had the next best result. Against the SiW-M dataset, the novel method achieved a mean EER of 6.72, plus or minus 5.66 percent, similarly outperforming the other models.
“Essentially, just fine-tuning a pre-trained vision transformer model for the PAD task was sufficient to achieve the state of the art performance,” the researchers conclude. “The proposed approach achieves the state of the art performance in unseen attack protocols of two publicly available datasets. In addition to excellent performance in unseen attacks, the pro-posed approach outperforms the state of the art methods in cross-datasets evaluations by a large margin, indicating the efficacy of the proposed approach in generalizing to both unseen attacks and domains.”