FRAN: Multi-scale Frequency–Spatial Residual Attention Network for General-Purpose AIGC Image Detection

The accelerating sophistication of deepfake technology erodes public trust and challenges the authenticity of digital media. In response, this paper presents a frequency-spatial sequential framework for deepfake detection that reframes forged content as multi-band, separable anomalies. The framework employs circular high-pass masks of varying radii in the frequency domain to partition the spectrum into distinct bands. Each band is processed in parallel through dual branches frequency amplitude and spatial reconstruction—yielding complementary representations. These band-wise features are arranged as a temporal sequence and fed into a bidirectional GRU, where an attention mechanism aggregates cross-scale evidence to accentuate minute forgery traces. To ensure computational efficiency, we retain only early convolutional layers as a local encoder and compress the resulting features via global average pooling. Extensive experiments on standard benchmarks confirm that the proposed method outperforms state-of-the-art approaches in deepfake detection accuracy while maintaining real-time inference capability.