FIND: An Internet wide Deepfake Detection Tool

The proliferation of deepfakes necessitates robust, multi- faceted detection mechanisms. This work introduces a synergistic framework that integrates two specialized pipelines to address the challenges of both video and image forgery detection [1]. For video deepfake detection (temporal analysis), a resource-efficient architecture combines a classical ResNet feature extractor with a Quantum Long Short Term Memory (QLSTM) network [1]. This quantum-enhanced model demonstrates up to a 99.16% reduction in recurrent layer parameters compared to classical LSTM, achieving performance superior to or comparable with its classical counterpart [1]. For image deepfake detection (frequency/spatial analysis), we adopt the WMamba module, which excels at exploiting subtle forgery artifacts by applying Dynamic Contour Convolution (DCConv) to hi- erarchical Wavelet representations [1]. WMamba achieves state-of- the-art (SOTA) performance in cross-dataset evaluation [1]. This dual approach confirms that specialized, multi-modal strategies are critical for resilient, computationally viable deepfake detection systems.