Multiple contexts and frequencies aggregation network for deepfake detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology is rapidly advancing, posing significant challenges for detection.
• Current methods often rely on domain-specific features, lacking generalizability.
• There is a need for robust deepfake detection models that capture intrinsic forgery characteristics.

Purpose of the Study:

• To propose an efficient and robust network, MkfaNet, for face forgery detection.
• To develop a backbone that learns generalizable spatial and frequency features for distinguishing real from fake samples.
• To improve deepfake detection performance across various datasets and manipulation types.

Main Methods:

• Designed MkfaNet, an efficient network incorporating two core modules: a Multi-Kernel Aggregator for spatial features and a Multi-Frequency Aggregator for frequency components.
• The Multi-Kernel Aggregator adaptively selects convolutional features to model subtle facial differences.
• The Multi-Frequency Aggregator adaptively re-weights high and low-frequency features for comprehensive analysis.

Main Results:

• MkfaNet achieved an Area Under the Curve (AUC) of 0.9591 in within-domain evaluations and 0.7963 in cross-domain evaluations on seven benchmarks.
• The proposed network outperformed several state-of-the-art deepfake detection methods.
• MkfaNet demonstrated high computational efficiency and enhanced robustness against diverse deepfake manipulations.

Conclusions:

• MkfaNet is an effective and efficient solution for deepfake detection.
• The network's ability to learn robust spatial and frequency priors contributes to its strong performance.
• The findings suggest MkfaNet offers improved generalization capabilities for identifying sophisticated face forgery techniques.