Abstract
The rapid development of face forgery technology has raised significant concerns regarding information security. While deepfake detection has proven to be an effective countermeasure, existing methods typically focus on binary classification accuracy, leaving predictive reliability largely unknown. Models often exhibit "overconfidence," assigning extremely high probabilities to incorrect predictions, especially when encountering unseen content. This paper introduces the concept of deepfake detection calibration using Packed-Ensembles (PE). By dividing a detector into multiple subnetworks, PE allows the model to focus on diverse forgery traces, providing reliable uncertainty estimation without incurring additional computational costs. Comprehensive experiments demonstrate that this method effectively mitigates overconfidence and enhances detection reliability while maintaining strong generalization performance.
The Challenge: Overconfidence in Deepfake Detection
Most research in deepfake detection focuses on binary accuracy—classifying an image as "real" or "fake." However, in high-stakes environments like legal proceedings or financial verification, knowing when a model is likely to be wrong is as crucial as the prediction itself.
Current detectors suffer from a phenomenon known as overconfidence. When a model encounters a forgery type it wasn't trained on, it often outputs an extreme probability (e.g., 0.99) even when its prediction is incorrect. This indicates a failure in model calibration—the alignment between a model’s predicted probability and the actual likelihood of correctness.
Methodology: Packed-Ensembles for Deepfake Detection (PE-DD)
To address these issues, the authors propose the PE-DD framework, which rests on two innovative pillars:
1. Packed-Ensembles (PE)
Traditional ensembles improve reliability by training multiple separate models, but they are computationally expensive. PE uses sparse grouped convolutions to "pack" multiple subnetworks into a single architecture. By dividing the network along channel dimensions, each subnetwork is forced to learn different forgery traces. This provides the diversity of an ensemble with the computational footprint of a single model.
2. High-Frequency Feature Injection
Deepfakes often leave subtle traces in the high-frequency domain that are invisible to the naked eye. The authors utilize SRM (Steganalysis-Rich Model) filters to extract edge and noise patterns. These features are concatenated with original image channels, helping the subnetworks detect imperceptible texture anomalies and artifacts.
Experimental Evaluation
The method was evaluated on the FaceForensics++ (FF++) dataset for training, with extensive testing on out-of-distribution (OOD) benchmarks including Celeb-DF v2, DFDC, Wild Deepfake (WDF), and AI-generated images (GenImage).
Key Results and Reliability
The primary metric for reliability is the Expected Calibration Error (ECE). A perfectly calibrated model would have an ECE of zero, where its confidence matches its accuracy exactly.
- Calibration Gains: On the FF++ dataset, PE-DD reduced the ECE from 0.085 (standard model) to 0.020, representing a massive improvement in reliability.
- Robustness to Noise: Under real-world "distribution shifts" like Gaussian noise or JPEG compression, PE-DD maintained significantly higher AUC and lower ECE than standard models, which typically become wildly overconfident when faced with noise.
- Efficiency: PE-DD achieved better calibration than traditional Deep Ensembles (DE) while using only 25% of the parameters and maintaining higher throughput (images per second).
Conclusion
PE-DD demonstrates that deepfake detectors do not have to sacrifice accuracy for reliability. By leveraging internal subnetworks and high-frequency features, the framework provides a computationally efficient way to estimate uncertainty and mitigate overconfidence. This shift toward calibrated detection is a vital step toward deploying trustworthy AI in sensitive security domains.
Summary updated