Continuous fake media detection: adapting deepfake detectors to new generative techniques

TL;DR

This paper tackles the rapid evolution of generative media by evaluating continual learning strategies for deepfake detection. It analyzes Knowledge Distillation (KD) and Elastic Weight Consolidation (EWC) on short and long fake-media sequences from the CDDB dataset and presents a lightweight CI/CD pipeline for continuous detector maintenance. The results show that continual learning can mitigate catastrophic forgetting and maintain performance across evolving tasks, with task similarity and arrival order playing critical roles; multi-task grouping further improves long-sequence robustness. The work demonstrates a practical pathway to deploy continuously updated deepfake detectors that adapt to new generation techniques while enabling data drift monitoring and retraining triggers.

Abstract

Generative techniques continue to evolve at an impressively high rate, driven by the hype about these technologies. This rapid advancement severely limits the application of deepfake detectors, which, despite numerous efforts by the scientific community, struggle to achieve sufficiently robust performance against the ever-changing content. To address these limitations, in this paper, we propose an analysis of two continuous learning techniques on a Short and a Long sequence of fake media. Both sequences include a complex and heterogeneous range of deepfakes generated from GANs, computer graphics techniques, and unknown sources. Our study shows that continual learning could be important in mitigating the need for generalizability. In fact, we show that, although with some limitations, continual learning methods help to maintain good performance across the entire training sequence. For these techniques to work in a sufficiently robust way, however, it is necessary that the tasks in the sequence share similarities. In fact, according to our experiments, the order and similarity of the tasks can affect the performance of the models over time. To address this problem, we show that it is possible to group tasks based on their similarity. This small measure allows for a significant improvement even in longer sequences. This result suggests that continual techniques can be combined with the most promising detection methods, allowing them to catch up with the latest generative techniques. In addition to this, we propose an overview of how this learning approach can be integrated into a deepfake detection pipeline for continuous integration and continuous deployment (CI/CD). This allows you to keep track of different funds, such as social networks, new generative tools, or third-party datasets, and through the integration of continuous learning, allows constant maintenance of the detectors.

Figures (7)

• Figure 1: Proposed CI/CD pipeline for deepfake detection. The data coming from different sources like new generative tools, social media or existing databases are analyzed by forensic experts for continual retraining of the system. Next, these data are used for continual learning and monitoring. The data drift distribution module rise an alert whenever it detects new input data distributions. The continaul learning methods analyzed in this paper are part of the MLOps CI/CD pipeline block in the figure.
• Figure 2: Zero-shot performance on the Long set. All the models are trained only on the GauGAN task and evaluated over the whole dataset.
• Figure 3: Elastic Weight Consolidation average accuracy at each task $t$ calculated over tasks $\{1,\dots,t\}$. of all backbones on the Easy set. The order of the tasks is the following: (1) GauGAN, (2) BigGAN, (3) CycleGAN, (4) IMLE, (5) FaceForensic++, (6) CRN, and (7) WildDeepfake.
• Figure 4: Knowledge distillation average accuracy at each task $t$ calculated over tasks $\{1,\dots,t\}$. of all backbones on the Easy set. The order of the tasks is the following: (1) GauGAN, (2) BigGAN, (3) CycleGAN, (4) IMLE, (5) FaceForensic++, (6) CRN, and (7) WildDeepfake.
• Figure 5: The average accuracy of Resnet-50 trained with KD on the full Easy set (Figure \ref{['fig:short']}) and without WildDeepfake (Figure \ref{['fig:short-wd']}). Some datasets seem to heavily afflict performance in the continuous learning context.