{"id":"bdea9683-9f13-41d3-b733-95040c750b5e","arxiv_id":"2608.01661","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FairForensics combines expression-aware visual features, demographic-aware language alignment, and prototype-based fairness losses to improve cross-dataset deepfake detection and demographic fairness.","lead":"This paper presents FairForensics, a vision-language model that uses facial expression features and demographic text prompts to improve deepfake detection accuracy and demographic fairness. It also introduces a balanced demographic benchmark built from the GenFace dataset and reports gains over re-implemented baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The visual forgery amplification loss (Eq. 7-8) is instance-contrastive with each sample as its own positive and all other samples as negatives, with no authenticity labels; it cannot amplify forgery-specific features, so a named mechanism behind the SOTA claim is unsupported.","rationale":"The reader's verdict is REJECT, and my stress-test does not change that conclusion, so the verdict is UNCHANGED. I agree with the reader's list of misdescribed mechanisms, but I would put the VFA loss first among them. It is the most load-bearing because it is a named component, it is claimed to perform a specific function ('visual forgery amplification'), and the equation given in the paper directly contradicts that claim. The 308-token text-encoder issue is serious but could plausibly be resolved by extending positional embeddings; the IAPP identity-scoring-head issue could in principle be resolved by an implicit training signal. The VFA issue is visible purely from Eq. 7-8 and cannot be fixed by a small implementation clarification without changing the loss itself. The balanced benchmark and the broad experimental effort are real contributions, and the paper should receive credit for those. But the central claim of state-of-the-art generalization and fairness rests on components that, as written, do not behave as described. A concrete re-implementation of Eq. 7-8 and the Table V ablation would determine whether this concern is fatal to the causal narrative or merely a typographical error with an easily corrected formulation.","tokens_in":24463,"tokens_out":8484,"duration_ms":81241,"concrete_test":"Implement exactly Eq. 7-8 and train on the balanced GenFace split; monitor the mean cosine similarity within real, within fake, and between real/fake for I_v. If L_vfa is forgery-amplifying, within-class similarity should increase relative to between-class similarity; as written, all pairwise similarities are repelled equally, so the loss will not change the class-conditional gap. Then replace L_vfa with a generic feature-spreading regularizer and re-run the ablation in Table V: if the DFDC/FF++ gains are reproduced, the reported benefit cannot be attributed to the stated forgery-amplification mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D defines L_vfa as a contrastive loss over 'vision-vision feature pairs {(I^u_v, I^u_v)}'. In Eq. 7 the similarity is computed between I^u_v and I^v_v over batch indices u and v, and Eq. 8 uses one-hot labels y_pa with y^uu_pa = 1 and y^uv_pa = 0 for u != v. Thus the positive pair for anchor u is the feature vector of the same image itself, not a different view, a same-class sample, or a manipulated counterpart. There is no authenticity label in the loss at all. Minimizing this loss spreads all samples in the batch in feature space, repelling real-real pairs exactly as strongly as real-fake pairs. The paper's claim that this loss 'amplifies forgery features' and 'pull[s] positive pairs together while pushing negative ones away' is therefore not realized by the equation. Since L_vfa is listed as a named component (VFA) and its removal changes DFDC AUC by 3.79% in Table V, the causal story for this gain is missing: the gain could be a generic feature-dispersion effect rather than expression-aware forgery amplification. This is not an implementation detail; it is a mismatch between the central method description and the equation actually provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FairForensics, a vision-language framework for fair and generalizable deepfake detection on a demographically balanced benchmark constructed from GenFace. It combines an expression encoder built on POSTER, an expression-perceptual visual encoder with an identity-aware patch perturbation (IAPP) module, a demographic-guided language encoder (DGLE) that builds text prompts over authenticity, sex, age, and race, and a population-aware prototype fairness (PPF) objective. The authors report state-of-the-art cross-dataset AUC and fairness gaps on FF++, Celeb-DF, DFDC, and DF-1.0, with extensive ablations supporting the role of each component.","tokens_in":24712,"tokens_out":6252,"duration_ms":54395,"significance":"If validated, the balanced benchmark itself would be a useful resource for the fairness community, and the cross-dataset results suggest that demographic-aware language prompts and balanced training distributions can improve fairness without sacrificing detection performance. The paper ships no code or checkpoints at submission, and several named mechanisms conflict with the equations provided, so the current evidence for the specific design choices is incomplete. The ablation suite is broad (Tables IV-VIII) and the use of genuinely held-out public datasets for the headline claim is a real strength.","major_comments":[{"comment":"The VFA loss is claimed to 'amplify forgery features' and to 'pull positive pairs together while pushing negative ones away,' but the positive pair for each anchor is the feature of the same image itself. The one-hot label y_pa in Eq. (8) sets y^uu_pa = 1 and y^uv_pa = 0 for u ≠ v, so Eq. (7) is an instance-discrimination softmax over the batch, not a forgery-amplification objective. No authenticity labels enter the loss, so real-real and real-fake pairs are pushed apart with equal strength. Table V ascribes a 3.79% AUC gain on DFDC to L_vfa, but a generic feature-spreading effect could produce that gain; the claimed expression-aware amplification mechanism is not realized by the published equations. The authors should either reformulate the loss with genuinely forgery-relevant positives (e.g., same-class or expression-matched pairs) or remove the unsupported mechanistic claim.","section":"III-D, Eq. (7)-(8)"},{"comment":"T_emb is defined as being in R^{b x 308 x s} and is fed to the text encoder with a positional embedding P_t in R^{b x 308 x s}. The paper cites CLIP [14] as the text encoder and tokenizer; standard CLIP text transformers have a fixed 77-token context length with learned positional embeddings of size 77. No modification is described for handling 308 tokens, so as written the demographic language branch cannot be executed. Please specify the architecture change (e.g., extended positional embeddings, per-sentence encoding followed by pooling, or a different text encoder) and update the parameter counts in Table III accordingly.","section":"III-C, DGLE and text encoder"},{"comment":"The identity sensitivity scores S = sigmoid(f_id(I_pat_2)) are produced by a two-layer MLP, but no identity labels, identity classification loss, or identity-contrastive objective is defined anywhere in the paper. Without any supervision linking f_id's output to identity, there is no reason the top-k selected patches correspond to identity-responsive regions; the scores could be driven by forgery cues or arbitrary image statistics. The claim that IAPP 'suppresses identity-related shortcut learning' is therefore unsupported. Add an identity-aware training signal or provide direct evidence (e.g., correlation with identity classification accuracy on the selected patches).","section":"III-B, Eq. (4) and IAPP"},{"comment":"The balanced GenFace benchmark is a re-split of the authors' own dataset with FairFace annotations, and the baseline methods are re-implemented by the authors using 'default configurations' with no released code. Since the protocol is new and baselines are not official checkpoints, the within-dataset comparisons in Table II and the SOTA claim in the abstract cannot be independently verified without releasing the evaluation code, the split indices, and the trained models. At minimum, release the benchmark construction code and the exact train/validation/test identity-disjoint splits; ideally provide official-baseline results on the same splits.","section":"IV-A, benchmark and baselines"}],"minor_comments":[{"comment":"Reference [24] is cited for POSTER in Section III-B but [24] is the ETH-XGaze gaze-estimation paper; the correct POSTER citation is [2]. The reference list also lacks the actual POSTER entry under [24].","section":"References / Section III-B"},{"comment":"The text says that inaccurate demographic prompts make fairness scores 'improve noticeably,' while the table shows F_FPR and F_MEO increase (worsen) for both MFCLIP and FairForensics (e.g., Ours w/ASAR* has F_FPR 10.45 vs Ours w/ASAR† with 16.24 on FF++). Please reword to match the reported numbers.","section":"IV-C, Table IV discussion"},{"comment":"The row label 'Ours w/ASA' appears to be a typo for 'Ours w/ASG', given the column headings A, S, G, and R.","section":"Table IV"},{"comment":"There are punctuation typos that impede reading: 'Params 421.587' should likely be '421.59', and the CViT row shows '49,89' with a comma instead of a decimal point.","section":"Table III"},{"comment":"The text 'senmatic signals' contains a typo for 'semantic signals'; also the notation I^r_vA and I^f_vA in the L_vlc explanation is not defined in the notation list.","section":"III-D, DSAD description"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a timely topic and a broad empirical study, but the method section contains equation-to-text inconsistencies that a careful reader will notice immediately. The VFA loss, the 308-token text encoder issue, and the unsupported IAPP supervision are load-bearing for the claimed mechanisms. I would ask the authors to address these points and to release benchmark and baseline code before a revised submission is considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the thing to know about FairForensics is that it gets real cross-dataset gains over fair-deepfake baselines, but two of its named mechanisms are not what they claim to be as written. The balanced GenFace split and the prototype fairness loss are worth taking seriously; the VFA loss and IAPP identity scoring are not.\n\nWhat is actually new: the balanced demographic benchmark protocol (equalizing eight intersectional subgroups in GenFace and testing under balanced/unbalanced shifts) is a practical contribution. The cross-dataset numbers are noticeably above MFCLIP and other VLMs — 63.04, 62.35, 59.88, 62.75 AUC on FF++, Celeb-DF, DFDC, DF-1.0 versus roughly 51, 54, 51, 50 for MFCLIP. The PPF loss (margin plus subgroup-prototype alignment) is clearly formulated and its ablations suggest it helps both fairness and detection. The expression-FID observation is useful as a heuristic for choosing an expression backbone, even if calling it a novel finding overstates it — any feature extractor will show some real/fake distribution shift.\n\nThe soft spots are not minor. First, the VFA loss (Eq. 7–8) is instance-contrastive: the only positive pair for a sample is itself, and there is no authenticity label in the loss. It cannot 'amplify forgery features' as the paper claims; it spreads all samples in the batch. The 3.79% DFDC drop when removing it (Table V) therefore has no stated causal mechanism. Second, the IAPP identity scoring head is a sigmoid on a two-layer MLP with no identity supervision; calling it 'identity sensitivity' is not supported. It may still work as a regularizer, but the name and the claim are unjustified. Third, the demographic-guided text encoder concatenates four 77-token sentences into a 308-token sequence and adds a 308-position embedding, but the paper never explains how a standard CLIP encoder with 77 positional embeddings handles this. That is a missing technical detail that could be fixed, but it blocks reproduction. Finally, all experiments are single runs with no error bars, and code/data are only promised. Given the baseline re-implementations are also not released, the apparent 10-point gains deserve more scrutiny.\n\nWho is this for? Researchers in fair deepfake detection and cross-dataset generalization. The balanced benchmark idea and the PPF objective are valuable. But as written, the paper overclaims. I would send it to peer review — the empirical picture is strong enough to merit referee time — with the expectation of major revision: rework or properly reframe VFA, add identity supervision to IAPP or drop the name, specify the text-encoder modification, release the benchmark and code, and report variances. If the authors fix those, this could be a solid contribution.","headline":"Real cross-dataset gains and a useful balanced fairness benchmark, but the VFA loss and IAPP mechanism are misdescribed as written, and the 308-token text encoder detail is unresolved.","tokens_in":25265,"tokens_out":4655,"would_cite":false,"duration_ms":38417,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FairForensics pairs expression cues with demographic language prompts to make deepfake detection both more generalizable and fairer across demographic groups.","keywords":["fair deepfake detection","vision-language model","demographic fairness","facial expression forgery","identity-aware patch perturbation","prototype fairness","cross-dataset generalization","balanced demographic benchmark"],"falsifier":"Open the released code and feed a 308-token input to the text encoder: if the encoder is the standard CLIP text tower with 77-token positional embeddings, it will either error, truncate, or silently repeat positions, and inspecting whether the positional embedding table was interpolated or replaced would settle whether the demographic language branch works as described. A second check is to train the full model with the demographic language branch removed and compare cross-dataset AUC and fairness on Celeb-DF; the paper's own ablations predict a large drop, so observing no drop would cast doubt on the mechanism.","tokens_in":24198,"feed_emoji":"🎭","tokens_out":7959,"duration_ms":69817,"temperature":0.7,"pith_summary":"This paper tries to establish that deepfake detection can be made both more accurate on unseen forgeries and fairer across demographic groups by combining two signals that prior detectors mostly ignore: high-level facial-expression differences between real and fake faces, and language prompts that name a face's demographic attributes. It introduces FairForensics, whose central claim is that expression-guided forgery features, identity-suppressing patch perturbations, demographic-aware text supervision, and prototype-level fairness regularization together produce representations that are discriminative for forgery yet invariant to identity and demographics. To support this, the paper builds a demographically balanced benchmark from GenFace with eight intersectional subgroups (sex, age, and race) and evaluates under balanced and imbalanced training as well as cross-dataset transfer to FF++, Celeb-DF, DFDC, and DF-1.0. If correct, the work matters because it offers a concrete recipe for stopping detectors from latching onto who a face is rather than whether it was manipulated.","feed_headline":"Expression plus demographic language yields fairer deepfake detection","feed_subtitle":"The model trains on balanced face groups and reports higher AUC with lower demographic fairness gaps on four unseen datasets.","key_machinery":"The machinery is a multi-module training pipeline. The expression encoder (EE) takes the 8-dimensional embedding from a frozen POSTER expression estimator and adapts it through a trainable adapter and forgery projector, turning expression vectors into expression-guided forgery features supervised by their own real/fake loss. The expression-perceptual visual encoder (EPVE) is a vision transformer whose blocks contain an expression injector (EI): the class token queries the expression features in parallel attention, so global appearance tokens learn to absorb expression forgery cues. After the first block, an identity-aware patch perturbation module (IAPP) scores each patch token for identity sensitivity, then during training adds Gaussian noise scaled by that score and drops a subset of the top-k identity-responsive patches, forcing the model to rely on non-identity regions. The demographic-guided language encoder (DGLE) builds four prompts (authenticity, sex, age, race), tokenizes each to 77 tokens, concatenates them into a 308-token sequence, and encodes it with a text transformer to obtain population-aware language embeddings aligned to visual features through a contrastive loss. Finally, the population-aware prototype fairness objective (PPF) computes class-conditional subgroup prototypes and applies a margin loss separating real and fake prototypes across groups plus an alignment loss pulling same-class prototypes together; these are combined with visual forgery, expression forgery, vision-vision contrastive, and language discrimination losses in a single objective.","core_discovery":"The paper claims that deepfake detection can be simultaneously generalized and made fair by treating facial expression as a structured forgery signal and demographic attributes as language-level supervision. FairForensics builds on the observation that real and fake faces differ in their expression-embedding distributions, and uses that difference to inject expression-aware forgery cues into a vision transformer while an identity-aware patch perturbation suppresses person-specific shortcuts. On the demographic side, text prompts describing authenticity, sex, age, and race are encoded together and aligned with visual features, and a prototype-level objective pulls same-class demographic subgroups together while pushing real and fake prototypes apart. On its balanced GenFace benchmark and cross-dataset tests, the paper reports AUC of 63.04% on FF++, 62.35% on Celeb-DF, 59.88% on DFDC, and 62.75% on DF-1.0, with lower false-positive-rate and equalized-odds gaps than the compared detectors, and it states that this is state of the art in both generalization and fairness.","pith_inferences":["If the expression-distribution signal is as general as the paper argues, any frozen expression estimator whose real/fake embeddings are well separated could serve as the forgery prior, so a cheap test is to swap POSTER for another estimator and measure cross-dataset AUC.","The paper's Table IV shows that inaccurate race prompts hurt fairness, implying that demographic annotation noise is a sensitivity point; a natural extension is evaluating the framework under noisy or missing demographic labels.","The 308-token text-encoder question is testable in the released code: if the standard 77-token CLIP position embeddings are not extended, the demographic language branch behaves differently from the description, and the fairness gains attributed to it would need re-attribution.","The benchmark's demographic bins (0-29 versus 30 and over, white versus non-white, male versus female) are coarse, and the paper itself notes residual bias under complex intersections, so extending the prototype and language machinery to finer-grained or continuous attributes is a natural next step."],"forward_implications":["Balanced demographic training data alone lowers fairness gaps and improves cross-dataset AUC relative to imbalanced training for the same detector.","Adding sex, age, and race text prompts on top of authenticity prompts improves both detection AUC and demographic fairness on unseen datasets.","Perturbing identity-sensitive patches during training reduces identity-shortcut reliance and improves fairness on cross-dataset tests with only a small parameter increase.","Removing either the prototype margin loss or the subgroup alignment loss degrades cross-dataset AUC and fairness, so both components are needed.","The model is more robust to unseen image distortions such as Gaussian noise, blur, and compression than the compared detectors."],"supporting_citations":[{"why":"Supplies the GenFace benchmark and protocol from which the balanced demographic training, validation, and test splits are constructed.","marker":"[17]"},{"why":"Provides the frozen POSTER expression estimator whose embeddings define the expression forgery prior used throughout the method.","marker":"[24]"},{"why":"Provides the CLIP text encoder and vision-language contrastive objective used by the demographic-guided language encoder and the VLC loss.","marker":"[14]"},{"why":"Serves as the vision-language deepfake detection baseline that FairForensics extends and compares against in prompt and fairness ablations.","marker":"[1]"},{"why":"Represents the fairness-enhanced deepfake detector baseline whose cross-domain fairness results FairForensics claims to surpass.","marker":"[11]"},{"why":"Documents identity leakage in deepfake detectors, motivating the identity-aware patch perturbation module.","marker":"[15]"},{"why":"Supplies the FairFace automatic sex, age, and race annotations used to label the intersectional demographic subgroups.","marker":"[27]"},{"why":"Provides ForensicAdapter, a CLIP-based forgery detector baseline used in cross-dataset comparisons.","marker":"[20]"}],"fun_headline_variants":["Expression and demographic cues yield fairer deepfake detection","Balanced demographics plus expression clues improve deepfake detection","Vision-language model for fair and general deepfake detection","FairForensics: seeing expressions, parsing demographics","Demographic-aware vision-language improves deepfake fairness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise of the demographic language branch is that a CLIP-style text encoder can ingest the concatenated 308-token prompt, yet the paper never specifies how the encoder's 77-token positional embeddings are extended, so if that step is unstated the branch cannot operate as described.","fun_headline_variants_meta":{"raw":{"variants":["Expression and demographic cues yield fairer deepfake detection","Balanced demographics plus expression clues improve deepfake detection","Vision-language model for fair and general deepfake detection","FairForensics: seeing expressions, parsing demographics","Demographic-aware vision-language improves deepfake fairness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1668,"prompt_tokens":1010,"completion_tokens":658,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":584}},"tokens_in":626,"tokens_out":658,"duration_ms":5616,"temperature":1.0,"reasoning_tokens":584,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:06:07.349954+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the released code and feed a 308-token input to the text encoder: if the encoder is the standard CLIP text tower with 77-token positional embeddings, it will either error, truncate, or silently repeat positions, and inspecting whether the positional embedding table was interpolated or replaced would settle whether the demographic language branch works as described. A second check is to train the full model with the demographic language branch removed and compare cross-dataset AUC and fairness on Celeb-DF; the paper's own ablations predict a large drop, so observing no drop would cast doubt on the mechanism.","supporting_citations":[{"cited_title":"Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation,","cited_arxiv_id":null,"evidence_quote":"Provides the frozen POSTER expression estimator whose embeddings define the expression forgery prior used throughout the method."},{"cited_title":"Preserving fairness generalization in deepfake detection,","cited_arxiv_id":null,"evidence_quote":"Represents the fairness-enhanced deepfake detector baseline whose cross-domain fairness results FairForensics claims to surpass."},{"cited_title":"Implicit identity leakage: The stumbling block to improving deepfake detection generalization,","cited_arxiv_id":null,"evidence_quote":"Documents identity leakage in deepfake detectors, motivating the identity-aware patch perturbation module."}],"review_version":2}