Exploring ChatGPT for Face Presentation Attack Detection in Zero and Few-Shot in-Context Learning

Abstract: This study highlights the potential of ChatGPT (specifically GPT-4o) as a competitive alternative for Face Presentation Attack Detection (PAD), outperforming several PAD models, including commercial solutions, in specific scenarios. Our results show that GPT-4o demonstrates high consistency, particularly in few-shot in-context learning, where its performance improves as more examples are provided (reference data). We also observe that detailed prompts enable the model to provide scores reliably, a behavior not observed with concise prompts. Additionally, explanation-seeking prompts slightly enhance the model's performance by improving its interpretability. Remarkably, the model exhibits emergent reasoning capabilities, correctly predicting the attack type (print or replay) with high accuracy in few-shot scenarios, despite not being explicitly instructed to classify attack types. Despite these strengths, GPT-4o faces challenges in zero-shot tasks, where its performance is limited compared to specialized PAD systems. Experiments were conducted on a subset of the SOTERIA dataset, ensuring compliance with data privacy regulations by using only data from consenting individuals. These findings underscore GPT-4o's promise in PAD applications, laying the groundwork for future research to address broader data privacy concerns and improve cross-dataset generalization. Code available here: https://gitlab.idiap.ch/bob/bob.paper.wacv2025_chatgpt_face_pad

• The paper demonstrates that GPT-4o achieves 90.91% accuracy for print attacks and 98.81% for replay attacks in a two-shot learning scenario.
• The paper leverages detailed in-context prompting and explanation requests to significantly reduce Failure-to-Acquire incidents and improve interpretability.
• The paper highlights the promise of multimodal language models in biometric security, paving the way for scalable and robust presentation attack detection systems.

Exploring ChatGPT for Face Presentation Attack Detection in Zero and Few-Shot in-Context Learning

The paper presents a thorough investigation into the application of the LLM GPT-4o, a variant of ChatGPT, for Face Presentation Attack Detection (PAD), specifically focusing on zero and few-shot in-context learning scenarios. Face PAD is a critical area in the field of biometric security, dealing with preventing unauthorized access through face recognition systems using spoofing techniques such as printed images or replayed videos.

Key Findings and Results

The research highlights that GPT-4o shows notable potential for PAD tasks, especially in few-shot learning contexts. A remarkable aspect of the study is the model's emergent reasoning ability, enabling it to predict the type of presentation attack—print or replay—with significant accuracy, even when not explicitly trained for this classification task. In a few-shot learning scenario with two examples, GPT-4o achieved an accuracy of 90.91% for print attacks and 98.81% for replay attacks. This performance is highly competitive when compared against specialized PAD systems.

In terms of traditional PAD metrics, the Average Classification Error Rate (ACER) for GPT-4o in a two-shot scenario was 2.7%, which is close to the performance of state-of-the-art systems like DeepPixBis trained on the SOTERIA dataset that reported an ACER of 2.0%. This result is achieved by leveraging context and interpreting detailed prompts, which facilitate the model's ability to provide consistent and accurate authenticity scores in the presence of reference data.

Methodological Approach

The authors employed a meticulous methodological framework, assessing the model's performance through various configurations: zero-shot, one-shot, and two-shot scenarios. Each of these configurations provides a different number of reference examples to the model to potentially enhance its performance. The study extensively uses the SOTERIA dataset, respecting data privacy regulations by limiting experiments to data from consenting individuals. Innovative prompting strategies were explored, and the impact of explanation requests on model output was carefully analyzed.

Using detailed prompts was critical to reducing Failure-to-Acquire (FTA) incidents, which occur when the model fails to generate an output. Short prompts led to a high FTA rate of 80.34%, which was significantly reduced by using detailed prompts. Additionally, explainability in model outputs enhanced interpretability with a slight performance improvement, demonstrating the usefulness of transparency in machine learning models.

Implications and Future Directions

The findings suggest that GPT-4o can serve as a versatile tool for biometric security applications, offering a flexible, context-based approach to face PAD without extensive domain-specific training. This shifts the paradigm from traditional vision-centric models to multimodal approaches that integrate language understanding and reasoning capabilities.

The study opens avenues for further research, particularly in applying LLMs to broader, cross-dataset scenarios and addressing challenges of generalization in PAD systems. Future research might focus on improving the scalability and robustness of these models in varied real-world applications while maintaining ethical and privacy standards. Further exploration into locally deploying such models could mitigate privacy concerns associated with cloud-based model deployments. Additionally, the evolution of multimodal LLMs could see their application extend beyond PAD into other areas of biometric security and computer vision tasks.

In conclusion, the paper illustrates that advancements in LLMs like GPT-4o provide a promising alternative for tackling sophisticated tasks in biometric PAD, leveraging their inherent reasoning capabilities and flexibility in handling low-resource scenarios. These properties render them highly adaptable to evolving security needs and multidomain challenges.