Source-linked AI summary
DefakeHop: A Light-Weight High-Performance Deepfake Detector
Hong-Shuo Chen, Mozhdeh Rouhsedaghat, Hamza Ghani, Shuowen Hu, Suya You, C. -C. Jay Kuo
TL;DR
Deepfake volume and quality have increased detection demands, while deep-learning detectors can be large and costly to train. DefakeHop combines SSL-based face features, feature distillation, and region/frame ensembling, achieving best performance on UADFV, Celeb-DF v1, and Celeb-DF v2.
Problem
Deepfake videos are becoming more numerous and harder to distinguish from real videos, while deep-learning detectors are typically large and expensive to train.
Source: §1 · p. 1Method
DefakeHop uses SSL-based PixelHop++ features from face-image regions, c/w Saab transforms, feature distillation, and ensembles across regions and frames.
Source: §1 · p. 1, §3.2 · p. 2, §3.2.1 · p. 3Results
DefakeHop achieves the best performance on UADFV, Celeb-DF v1, and Celeb-DF v2; its FF++/DF AUC is 2.15% below the best reported, with an upper-bound model size of 42,845 parameters.
Source: §4.1 · p. 4, §4.2 · p. 5Takeaways & Limitations
The authors conclude that DefakeHop combines smaller model size, fast training, high detection AUC, and fewer training samples.
Source: §5 · p. 5
Abstract
from arXiv · showhide
A light-weight high-performance Deepfake detection method, called DefakeHop, is proposed in this work. State-of-the-art Deepfake detection methods are built upon deep neural networks. DefakeHop extracts features automatically using the successive subspace learning (SSL) principle from various parts of face images. The features are extracted by c/w Saab transform and further processed by our feature distillation module using spatial dimension reduction and soft classification for each channel to get a more concise description of the face. Extensive experiments are conducted to demonstrate the effectiveness of the proposed DefakeHop method. With a small model size of 42,845 parameters, DefakeHop achieves state-of-the-art performance with the area under the ROC curve (AUC) of 100%, 94.95%, and 90.56% on UADFV, Celeb-DF v1 and Celeb-DF v2 datasets, respectively.
1. INTRODUCTION
Deepfake videos are growing more numerous and harder for people to distinguish from real videos, motivating automatic detection. DefakeHop offers a non-DL approach that combines feature extraction, distillation, and ensemble classification.
1. INTRODUCTION
- Deepfake video volume and quality are rising, while people increasingly struggle to distinguish fakes without side-by-side comparison.The introduction cites 7,964 clips online at the beginning of 2019 and 14,678 nine months later.Source: §1 · p. 1
- Deep-learning detectors are typically large and expensive to train, whereas non-DL methods usually perform worse.The paper notes models with hundreds of thousands or millions of parameters and costly training.Source: §1 · p. 1
- DefakeHop combines PixelHop++ feature extraction, feature distillation, and ensemble classification for non-DL Deepfake detection.It extracts features from face-image regions using SSL-based PixelHop++ and distills them into a more concise face description.Source: §1 · p. 1
2. BACKGROUND REVIEW
The background reviews handcrafted-feature approaches and deep-learning detectors, including both image-focused CNNs and systems that combine spatial and temporal features.
2.1. Non-DL-based Methods
- Non-DL approaches use handcrafted cues such as head-pose/landmark discrepancies, individual facial patterns, or visual artifacts, then apply classifiers such as SVMs or logistic regression.The cited examples include a one-class SVM trained only on real videos of high-profile individuals.Source: §2.1 · pp. 1–2
2.2. DL-based Methods
- CNN-based detectors target image evidence, including face-warping artifacts, mesoscopic patterns, and differences among facial regions.The reviewed work includes VGG and ResNet architectures, a 27,977-parameter network, and Xception applied to facial regions.Source: §2.2 · p. 2
- Integrated CNN/RNN approaches combine spatial and temporal features for Deepfake detection.Examples pair DenseNet with a bidirectional RNN or process InceptionV3 features with an LSTM before shallow detection.Source: §2.2 · p. 2
3. DEFAKEHOP METHOD
DefakeHop processes face images through PixelHop++, feature distillation, and ensemble classification. Its design reduces spatial and channel redundancy before combining evidence across facial regions and video frames.
3. DEFAKEHOP METHOD
- DefakeHop comprises PixelHop++, feature distillation, and ensemble classification after face-image preprocessing.These modules form the method's stated pipeline.Source: §3 · p. 2
3.1. Face Image Preprocessing
- Face preprocessing crops faces from video frames, aligns and normalizes them, then supplies consistent inputs to later modules.The procedure also samples frames, extracts landmarks, resizes and rotates faces, and crops patches from regions such as the eyes and mouth.Source: §3.1 · p. 2
3.2. PixelHop++ Module
- PixelHop++ applies three cascaded 3×3, stride-one units to local blocks of 32×32 color face patches, extracting rich discriminant features.Its spectral filtering uses c/w Saab transforms, while 2×2-to-1×1 max pooling reduces spatial redundancy.Source: §3.2 · p. 2
3.2.1. Channel-wise (c/w) Saab Transform
- The c/w Saab transform separates local mean and frequency components, then routes channels by frequency: high-frequency responses are discarded, mid-frequency responses retained, and low-frequency responses transformed further.Channel separability reduces the Saab model size without performance degradation; deeper hops trade spatial detail for a broader view.Source: §3.2.1 · pp. 2–3
3.3. Feature Distillation Module
- Feature distillation applies spatial PCA and per-channel soft classification to compact PixelHop++ outputs before a final face-patch classifier.It retains top PCA components containing about 90% of input energy; XGBoost soft classifiers use maximum depth one, and their channel probabilities form a K11-dimensional description.Source: §3.3 · p. 3
3.4. Ensemble Classification Module
- Ensemble classification combines soft decisions across facial regions and adjacent frames, then averages frame probabilities for a video-level fake probability.The reported regions are the left eye, right eye, and mouth; each frame is combined with three frames before and after it.Source: §3.4 · pp. 3–4
4. EXPERIMENTS
DefakeHop is evaluated against other detection methods across four datasets, with experiments also examining facial-region ensembles, video quality, model size, and training-data needs. The reported benchmark results are best on three datasets and within 2.15% of the best reported AUC on FF++/DF.
4. EXPERIMENTS
- The evaluation compares AUC on UADFV, FF++, Celeb-DF v1, and Celeb-DF v2, reporting both frame- and video-level results for DefakeHop.UADFV uses a random 80/20 train/test split; FF++ and Celeb-DF provide test sets, and the frame-level evaluation uses key frames only.Source: §4 · p. 4, §4.1 · p. 4
4.1. Detection Performance
- Ensembling facial regions boosts AUC by up to 5%, with different regions contributing strengths across faces.The ensemble gives the best result among the evaluated facial-region configurations.Source: Table 1 · p. 4, §4.1 · p. 4
- DefakeHop achieves the best performance on UADFV, Celeb-DF v1, and Celeb-DF v2; its FF++/DF AUC is 2.15% below the best reported result.It outperforms other non-DL methods and is competitive with DL methods [1] [6] [5].Source: §4.1 · p. 4, Table 2 · p. 5
- DefakeHop's performance degrades by 5% as video quality worsens, while the study reports reasonable handling of videos at different qualities.The experiments focus on compressed videos, which are challenging for Deepfake detection algorithms.Source: §4.1 · p. 4, Table 3 · p. 5
4.2. Model Size Computation
- The final model size is estimated at 42,845 parameters, an upper bound based on maximum XGBoost depth and channel count.The estimate accounts for PixelHop++ units, spatial PCAs, and channel-wise and ensemble XGBoosts.Source: §4.2 · pp. 4–5, Table 4 · p. 5
- About 85% AUC is reached with 250 training videos, and nearly 90% with 1,424 videos on Celeb-DF v2.These correspond to less than 5% and 23.6%, respectively, of the 6,011 training videos.Source: §4.2 · p. 5
5. CONCLUSION
The paper proposes DefakeHop as a light-weight, high-performance method for Deepfake detection. It reports a small model, fast training, high detection AUC, and fewer training samples as advantages, with experiments demonstrating detection performance.
5. CONCLUSION
- DefakeHop is proposed as a light-weight, high-performance method for Deepfake detection.The paper reports extensive experiments to demonstrate its detection performance.Source: §5 · p. 5
- The paper identifies smaller model size and a fast training procedure among DefakeHop's advantages.Source: §5 · p. 5
- The paper reports high detection AUC and fewer required training samples as advantages of DefakeHop.Fig. 5 plots AUC values as a function of training-video number.Source: §5 · p. 5, Figure 5 · p. 6