11dez2025
14:00 Master's Defense room 85 of IC2
Topic on
Lightweight, robust, and open-set facial anti-spoofing for mobile computing scenarios.
Student
Jadson Crislan Santos Costa
Advisor / Teacher
Anderson de Rezende Rocha
Brief summary
Biometric authentication, especially Facial Recognition (FRT), has become ubiquitous on mobile devices. However, its reliability is threatened by Presentation Attacks (PAs), such as the use of photos and videos, which aim to deceive systems. Presentation Attack Detection (PAD) seeks to mitigate this risk, but faces three main challenges in the mobile context: the need for computational efficiency (lightweight models), the ability to generalize to unknown attacks (open-set), and robustness to different domains, such as distinct cameras and lighting (cross-domain). Many state-of-the-art (SOTA) methods are computationally heavy or fail in these generalization scenarios. This dissertation proposes and evaluates a new lightweight (∼5M parameters), efficient, and robust PAD framework. The method uses modern architectures – a Transformer (DeiT-Tiny) and a CNN (ConvNeXtV2-Femto) – as feature extractors. We argue that pretraining with Self-Supervised Learning (SSL) methods from the Masked Image Modeling (MAE/FCMAE) family, which preserve spoofing artifacts (such as print textures or moiré patterns), is more suitable for PAD than contrastive methods that destroy these signals through color augmentations (e.g., color jitter). After pretraining, we performed supervised fine-tuning investigating two novel cosine-similarity-based loss formulations designed for the binary PAD scenario: (1) a Closed-Set approach (A1) that uses k-centers (k=10) and soft-labels (generated via Mixup/CutMix) to model the high intra-class variance of attacks; and (2) an Open-Set approach (A2) that uses a custom binary margin loss to enforce strict separation between bona fide samples and attacks. We exhaustively evaluated our framework on the RECOD-MPAD and Echoface-Spoof (RGB-only) datasets, comparing it to relevant baselines (ResNet-18, M³FASvisual, SSAN) in intra-dataset, cross-device, cross-attack, and cross-dataset protocols. The results demonstrate that, in intra-dataset scenarios, our methods achieve near-perfect performance (e.g., 0.04% ACER). More importantly, in generalization tasks (cross-device and cross-attack), our lightweight approaches, especially ConvNeXtV2 with both proposed losses, consistently outperform heavier baselines such as SSAN. Cross-dataset generalization proved to be the most difficult challenge for all methods, although our approaches remained competitive. This research demonstrates that it is feasible to create PAD models that are simultaneously lightweight, efficient, and robust for generalization, making them suitable for deployment on mobile devices.
Examination Board
Headlines:
| Anderson de Rezende Rocha | IC / UNICAMP |
| Giovani Chiachia | SAFFE |
| Hélio Pedrini | IC / UNICAMP |
Substitutes:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| David Menotti Gomes | DInf / UFPR |