Forgery Guided Learning (FGL) Overview

Updated 8 July 2026

Forgery Guided Learning is a paradigm that uses manipulation-specific signals (e.g., residual maps, discrepancy maps) to direct feature learning for improved forgery detection.
It integrates methods like cross-domain adaptation, weak supervision, and federated learning to enhance the robustness and generalization of detection systems.
Key strategies include reconstruction-based representation shaping, prompt-driven cues, and adaptive parameter updates, leading to high accuracy and precision in various forensic settings.

Forgery Guided Learning (FGL) is a forgery-analysis paradigm in which supervision, feature design, and optimization are driven by signals that reflect manipulation itself—such as residual maps, discrepancy maps, manipulation masks, frequency anomalies, prompts, or domain-difference signals—rather than relying only on generic real-versus-fake classification. In the current literature, the term is used explicitly for cross-domain deepfake adaptation in “Forgery Guided Learning Strategy with Dual Perception Network for Deepfake Cross-domain Detection” (Jia et al., 14 Aug 2025). A broader, technically consistent reading is that several adjacent methods instantiate the same idea implicitly by forcing models to learn from forgery-specific evidence: reconstruction residuals in federated face forgery detection (Liu et al., 2022), discrepancy-guided reconstruction for image forgery detection (Shi et al., 2023), weakly supervised forgery cue discovery in unpaired faces (Tian et al., 2024), prompt- and tool-guided visual-language systems (Chen et al., 16 Jul 2025), and weakly supervised audio forgery localization (Wu et al., 3 May 2025).

1. Conceptual scope

In the explicit formulation, FGL is a strategy that “captures the differential information between known and unknown forgery techniques,” allowing the detector to “dynamically adjust its learning process in real time” for cross-domain deepfake detection (Jia et al., 14 Aug 2025). More generally, the literature supports a wider definition: learning procedures in which the model architecture, auxiliary signals, and training objectives are organized around forgery cues—where the content is manipulated, how it was manipulated, and which representation best exposes that manipulation (Wu et al., 3 May 2025).

This scope spans several task families. In face forgery detection, the basic decision problem is binary classification of real versus forged faces under hybrid-domain and generalized-domain conditions, including unseen forgery types such as DeepFakes, FaceSwap, Face2Face, NeuralTextures, WildDeepfake, and DeeperForensics-style manipulations (Liu et al., 2022). In image forensics, the emphasis often shifts from classification to localization, with pixel-level masks for splicing, copy-move, inpainting, or diffusion-edited regions (Shi et al., 2023). In audio, the analogous problem is weakly supervised temporal forgery localization, where the objective is to identify forged segments of partially spoofed speech without frame-level annotation (Wu et al., 3 May 2025).

A central premise across these settings is that generic semantic features are insufficient. Several works argue that detectors overfit to dataset-specific textures, generator-specific artifacts, or semantic shortcuts, and therefore fail under cross-dataset or unseen-technique evaluation (Liu et al., 2022). This suggests that FGL is less a single architecture than a design principle: the detector should be constrained to use signals more causally linked to manipulation than ordinary image or audio semantics.

2. Principal guidance signals

The guidance signal in FGL is not fixed. Different systems expose forgery evidence through different intermediate objects, but the role of those objects is structurally similar: they bias the representation toward manipulation-sensitive information.

Mechanism	Guidance signal	Representative source
Residual-guided detection	$Res(x)=x-G(x)$	FedForgery (Liu et al., 2022)
Discrepancy-guided reconstruction	$\hat{R}_1=\|X-\hat{X}_1\|,\ \hat{R}_2=\|X-\hat{X}_2\|$	DisGRL (Shi et al., 2023)
Weakly supervised cue discovery	CAR maps and fused $\boldsymbol{a}_{\rm fus}$	FoCus (Tian et al., 2024)
Frequency-aware cue enhancement	Fine-grained DCT or wavelet/image-frequency adaptation	PEL (GU et al., 2021), FatFormer (Liu et al., 2023)
Prompt- and proposal-guided learning	Forgery-aware prompts, T-FAS, probability/location prompts	LOCO (Wu et al., 3 May 2025), MGFFD-VLM (Chen et al., 16 Jul 2025)
Parameter-update guidance	Learned $w_i,l_i$ or orthogonal forgery subspaces	FGL-DPNet (Jia et al., 14 Aug 2025), DevFD (Zhang et al., 23 Sep 2025)

Residual-guided approaches treat the reconstruction error itself as an artifact map. In FedForgery, the reconstruction network $G(\cdot)$ produces $\hat{x}=G(x)$ , and the classifier is trained on the residual $Res(x)=x-G(x)$ rather than the raw face, with a joint objective $L=\mu_1L_G+\mu_2L_{rec}+\mu_3L_{cls}$ (Liu et al., 2022). In DisGRL, discrepancy is generated twice through Double-Head Reconstruction, and the difference masks $\hat{R}_1,\hat{R}_2$ are fused with semantic and discrepancy features to form the detection signal (Shi et al., 2023). In FoCus, the guidance object is a weakly supervised manipulation map, obtained by combining RGB and Sobel cues through a Classification Attentive Regions Proposal module and a complementary learning module (Tian et al., 2024).

Prompt-based systems expose guidance in a different form. LOCO constructs forgery-aware prompts from utterance-level labels and learnable context, then uses proposal-generated pseudo labels and supervised semantic contrastive learning to refine frame-level audio representations (Wu et al., 3 May 2025). MGFFD-VLM turns the detector’s own outputs into prompts: classification probability becomes a “Probability Prompt,” and segmentation features become a “Location Prompt” for the LLM (Chen et al., 16 Jul 2025). A plausible implication is that FGL can operate either in pixel or feature space, or in token space, as long as the guidance channel remains forgery-specific.

3. Core architectures and objectives

A recurrent architectural pattern is reconstruction-driven representation shaping. FedForgery uses a VQ-VAE-like reconstruction network and a ResNet-50 classifier, with the classifier operating only on residual maps; the paper reports that residual features form more compact and distinct fake clusters than raw-image features in t-SNE analysis (Liu et al., 2022). DisGRL extends this logic by coupling a Discrepancy-Guided Encoder, a decoder, Double-Head Reconstruction, and a Discrepancy-Aggregation Detector, trained end-to-end with $L_{total}=L_{cls}+\lambda_1L_{r1}+\lambda_2L_{r2}+\lambda_3L_m$ (Shi et al., 2023). GIFL goes further by reframing localization as reconstruction of pristine content: with $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 0, $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 1, and $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 2, the model is trained by $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 3, thereby learning “universal features” from pristine regions rather than traces of specific forgeries (Zhao et al., 10 Apr 2025).

Another pattern is explicit cue discovery or localization under weak supervision. FoCus rejects paired real-forged comparison maps as supervision because they annotate all changed pixels, including compression and upsampling noise, and instead learns manipulation maps directly from image-level labels in unpaired data (Tian et al., 2024). Its CARP module is designed to expose multiple discriminative regions per class, while the complementary learning module performs token-wise RGB/Sobel competition through Gumbel-Softmax to produce fused tokens and the final manipulation map $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 4 (Tian et al., 2024). This makes the guidance signal itself a learned object rather than an externally computed pixel-difference heuristic.

Frequency-domain guidance is another major FGL axis. Progressive Enhancement Learning decomposes RGB images into a fine-grained frequency representation via patch-wise DCT in YCbCr space, then couples RGB and frequency branches through self-enhancement and mutual-enhancement modules, all trained with a single BCE objective (GU et al., 2021). FatFormer similarly argues that fixed CLIP-based linear probing lacks “forgery adaptation,” and inserts a forgery-aware adapter into the image encoder plus language-guided alignment with real/fake text prompts; with image- and frequency-domain adaptation together, it reports average detection accuracy of 98% on unseen GANs and 95% on unseen diffusion models (Liu et al., 2023). This suggests that frequency guidance is most effective when it is fused with spatial representation learning rather than treated as an isolated side channel.

4. Federated, personalized, continual, and adaptive variants

FGL is not confined to centralized training. FedForgery places residual-guided detection in a privacy-preserving federated setting, with each client training locally and the server aggregating parameters by $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 5 (Liu et al., 2022). The distinctive point is not the aggregation rule itself, which is standard weighted averaging, but that the federated object is a residual-guided detector whose clients observe heterogeneous artifact distributions. FedPR generalizes this personalized view by splitting each client model into a feature extractor $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 6, a personalized feature extractor $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 7, and a shared feature extractor $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 8, with losses $\hat{R}_1=|X-\hat{X}_1|,\ \hat{R}_2=|X-\hat{X}_2|$ 9, $\boldsymbol{a}_{\rm fus}$ 0, and an adversarial term $\boldsymbol{a}_{\rm fus}$ 1 combined as $\boldsymbol{a}_{\rm fus}$ 2 (Liu et al., 2024). In that formulation, forgery guidance is carried by client-specific and shared channel-statistics transformations.

Continual and developmental settings introduce a different kind of guidance: adaptation to newly emerging forgery types without forgetting. DevFD treats face forgery detection as continual learning and organizes LoRA experts into a shared Real-LoRA and multiple Fake-LoRAs, with new fake domains assigned orthogonal low-rank subspaces and trained under an integrated orthogonal loss $\boldsymbol{a}_{\rm fus}$ 3 (Zhang et al., 23 Sep 2025). The underlying assumption is that real faces are abundant and relatively stable, whereas fake faces evolve rapidly, so the model should preserve a shared real subspace while allocating new orthogonal subspaces to incremental forgery domains (Zhang et al., 23 Sep 2025). FGL-DPNet instead learns how to update parameters themselves: $\boldsymbol{a}_{\rm fus}$ 4, where $\boldsymbol{a}_{\rm fus}$ 5 and $\boldsymbol{a}_{\rm fus}$ 6 are generated from the current parameter and gradient state by a small adaptive network $\boldsymbol{a}_{\rm fus}$ 7 (Jia et al., 14 Aug 2025). Here, “guidance” resides in a learned optimizer that uses differential information between source-domain and target-domain forgeries.

A related training-time extension appears in simulated data generation. Agent4FaceForgery attributes the gap between benchmark accuracy and real-world efficacy to “ecological invalidity” of training data, then uses a multi-agent LLM framework with profile and memory modules to simulate forgery creation, social dissemination, and text-image inconsistency labels $\boldsymbol{a}_{\rm fus}$ 8, filtered by Adaptive Rejection Sampling $\boldsymbol{a}_{\rm fus}$ 9 (Lai et al., 16 Sep 2025). This can be read as an FGL curriculum in which the training set itself is guided by realistic forgery intent and detector difficulty.

5. Evaluation protocols and empirical evidence

The empirical literature evaluates FGL primarily under hybrid-domain, cross-dataset, cross-manipulation, weakly supervised, and few-shot unknown-technique protocols. FedForgery defines both a hybrid-domain forgery dataset and a generalized forgery dataset; on the hybrid-domain benchmark, centralized FedForgery reports 87.36% accuracy and 93.23 AUC, while the distributed federated version reports 85.55% accuracy and 91.12 AUC, preserving privacy with only a slight drop (Liu et al., 2022). Its ablation shows that replacing residual maps with raw images lowers hybrid-domain accuracy from 85.55% to 83.36%, and on DeeperForensics-1.0 the method achieves 95.21% accuracy on the standard-to-distorted protocol and 99.75% when both train and test are standard (Liu et al., 2022). FedPR reports 88.78% accuracy and 93.52% AUC on its Forgery Source Hybrid Dataset, and 97.29% accuracy on Deepforensics-1.0 in the federated setting (Liu et al., 2024).

DisGRL provides some of the clearest evidence that discrepancy-guided guidance improves unseen-forgery localization and classification. On WildDeepfake it reports 84.53 ACC and 93.27 AUC, and in cross-dataset evaluation trained on FF++(LQ) it reaches AUCs of 70.03 on Celeb-DF, 66.73 on WildDeepfake, and 70.89 on DFDC, all higher than RECCE in the reported comparisons (Shi et al., 2023). Its cross-manipulation average AUC on FF++(LQ) is 71.90, again the highest among the listed methods (Shi et al., 2023). FatFormer, trained on 4-class ProGAN data, reports average accuracy of 98% on unseen GANs and 95% on unseen diffusion models, explicitly framing this gain as the effect of forgery-aware adaptation and language-guided alignment (Liu et al., 2023).

Weakly supervised cue discovery and universal-feature learning show comparable trends. FoCus improves the evaluation model on FF++(HQ) to 96.43% accuracy and 99.15% AUC, and its maps improve several multi-task detectors in cross-dataset evaluation (Tian et al., 2024). GIFL reports a seen-forgery average of F1 $w_i,l_i$ 0, IoU $w_i,l_i$ 1, ACC $w_i,l_i$ 2, AUC $w_i,l_i$ 3, and an unseen-forgery average of F1 $w_i,l_i$ 4, IoU $w_i,l_i$ 5, ACC $w_i,l_i$ 6, AUC $w_i,l_i$ 7, outperforming trace-focused localization baselines on unseen manipulations (Zhao et al., 10 Apr 2025). In audio, LOCO extends the same principle beyond vision: under weak supervision it achieves EERs of 4.56% on HAD, 3.39% on LAV-DF, and 6.46% on AV-Deepfake-1M, with mAP values of 76.85%, 82.65%, and 18.89% respectively (Wu et al., 3 May 2025). The cross-modal consistency of these results suggests that FGL is best viewed as a modality-agnostic design principle rather than a face-only technique.

6. Interpretability, limitations, and terminological ambiguity

Interpretability is one of FGL’s strongest recurring themes. Residual-guided classifiers in FedForgery attend to semantic face regions such as eyes, mouth, and nose in Grad-CAM visualizations (Liu et al., 2022). DisGRL visualizes stronger discrepancy responses in forged regions and more precise Grad-CAM activation than baseline detectors (Shi et al., 2023). FoCus is explicitly motivated by the claim that ordinary attention maps often capture only partial or inaccurate forgery cues, and it presents manipulation maps as better supervision for downstream detectors (Tian et al., 2024). MGFFD-VLM turns classification and segmentation outputs into prompts so that a VLM can answer not only whether a face is forged but also why, while its text calibration loss explicitly supervises the authenticity word “real” or “fake” (Chen et al., 16 Jul 2025). ForgeryVCR makes this visual-centric logic explicit at the MLLM level: instead of text-centric chain-of-thought, it injects forensic tool outputs such as ELA, Noise Print++, FFT, and Zoom-In as visual intermediates, then optimizes tool usage with a tool utility reward (Wang et al., 15 Feb 2026).

The limitations are equally consistent. Residual-based methods depend on reconstruction quality: if reconstruction is too good or too poor, residuals become less informative, and background or margin noise can dominate (Liu et al., 2022). Personalized and federated methods incur communication overhead and require regular client participation with reasonably sized local datasets (Liu et al., 2024). Universal-feature localization can fail on small-scale forgeries, regular textures, complex content interference, or large low-detail regions (Zhao et al., 10 Apr 2025). VLM and MLLM systems inherit additional issues: MGFFD-VLM notes the need for better generalization to unseen manipulation methods (Chen et al., 16 Jul 2025), while ForgeryVCR argues that text-centric reasoning over low-level tampering traces leads to hallucination because linguistic modalities are poorly aligned with pixel-level forensic evidence (Wang et al., 15 Feb 2026). A plausible conclusion is that FGL improves robustness chiefly when the guidance signal is both manipulation-sensitive and operationally stable under domain shift.

Finally, the term itself is ambiguous. In multimedia forensics, “Forgery Guided Learning” denotes the family of strategies described above. Outside this area, however, FGL also denotes a successor to GL in ACL2 theorem proving, concerned with programmable rewriting and SAT-based proof rather than media forensics (Swords, 2020). Within forgery detection proper, another source of ambiguity is that many papers instantiate the idea without naming it. The literature therefore does not yet present a single canonical FGL formalism; instead, it offers a convergent body of methods in which forgery-specific residuals, discrepancies, masks, prompts, or adaptive updates guide representation learning more directly than label-only supervision.