Original title: Explainable Spatio-Temporal Attention Framework for GAN-Based Deepfake Detection
Abdullah Alhejaili, Muhammad Hatim Binsawad · September 2026
Fake videos made by computer programs, known as deepfakes, look very real and threaten trust online. Existing detection tools struggle to spot fakes made by different methods or in squished video files. They also fail to track changes across space and time well. To solve this, researchers built a new detection system. It uses an image network called XceptionNetA convolutional neural network architecture used here to extract spatial features from video frames. to examine details inside individual video frames. Then, it uses a sequence network called BiLSTMA recurrent network that reads a sequence both forward and backward to capture temporal context. and a temporal attention tool to track movements across time. This allows the system to focus on the most important frames. When tested on a major video collection, it achieved better accuracy at spotting fakes and naming the exact fake method used.
Imagine you are a referee watching a slow-motion replay to decide if a player dived. You can rewind and freeze frames, but the truth is in how the motion flows, not any single frame. Now imagine the league also asks you to point to the exact moment and body part that convinced you, not just give a verdict. In this paper the system watches video the same way: it reads each frame, tracks how the face moves across time, and then highlights which frames and regions led to its decision.
Deepfakes can be used to spread disinformation, commit fraud, or damage reputations. Most detectors fail when they meet a new manipulation method or a compressed video, and they give no reason for their verdict. This approach keeps working across several GAN-based manipulations and explains which parts of the video drove the decision, so humans can trust or audit the result.
The abstract doesn't say what the method fails on. It reports results only on FaceForensics++, so it does not establish performance on other datasets, real-world deepfakes outside the five listed manipulation types, or videos compressed in ways not represented in that dataset. It also gives no quantitative accuracy figures or comparisons to existing detectors.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Explainable Spatio-Temporal Attention Framework for GAN-Based Deepfake Detection — Abdullah Alhejaili, Muhammad Hatim Binsawad — Electronics — doi:10.3390/electronics15184343
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.