Kol-3-Freq: Frequency-Aware Video Deepfake Detection Using Kolmogorov-Arnold Networks

Abstract

Deepfake detection is a critical task today in multimedia forensics due to its unethical exploitation and serious consequences. Currently several deepfake detection techniques exist which involve synthetic human facial video detection, and most of which operate in the spatial domain While these methods are often very accurate (particularly, with respect to “in-dataset” analysis), none of them is foolproof, and are often susceptible to the sophistication of state-of-the-art deepfake generators. Very often, deepfakes generated by GANs, VAEs, or diffusion models are spatio-temporally coherent, but show prominent inconsistencies in the spectral domain. In this paper we propose <sc>Kol-3-Freq</sc>, a set of three neural deepfake detectors that detect frequency domain inconsistencies to classify real and synthetic videos. <sc>Kol-3-Freq</sc> models built on the novel Kolmogorov-Arnold Network, which makes the model architecture (<inline-formula> <tex-math notation="LaTeX">$\approx 2.181$ </tex-math></inline-formula> times) <italic>lightweight</italic> and more <italic>interpretable</italic> compared to the state-of-the-art, besides achieving a competitive performance improvement (<inline-formula> <tex-math notation="LaTeX">$\approx 1.4106\%$ </tex-math></inline-formula>) to its primitives when evaluated on standard deepfake datasets. <sc>Kol-3-Freq</sc> is further extended to four tiers–<sc>Light</sc>, <sc>Base</sc>, <sc>Large</sc>, and <sc>Huge</sc>, based on handling input-feature dimensionality of the underlying Kolmogorov-Arnold Network, making them <italic>adaptive</italic> and <italic>robust</italic>.