FakeChain: Forensics Benchmarks & Credentialing

• FakeChain is a multi-context research label covering deepfake benchmarking, multimodal forensic datasets, and blockchain-based credential verification.
• The deepfake benchmark tests detector robustness across 1-, 2-, and 3-step manipulations, illustrating a heavy reliance on final-stage artifacts.
• Its forensic dataset and blockchain application drive explainable reasoning and secure document validation, advancing anti-fraud methodologies.

FakeChain is a research label that appears in multiple technically distinct contexts in recent literature. In one usage, it denotes a large-scale benchmark for studying multi-step or hybrid deepfakes and the failure modes of current forensic detectors under sequential manipulation pipelines (Heo et al., 20 Sep 2025). In a second usage, it denotes a foundational multimodal dataset for “linguistic authenticity reasoning” in AI-generated image forensics, designed to connect “visual trace evidence” to structured authenticity judgments within the FakeScope framework (Li et al., 31 Mar 2025). In a third, implementation-oriented usage, “FakeChain” is used as a technical brief for a blockchain/IPFS-based academic credential verification system aligned with the Verifi-Chain prototype, where verified certificates are stored in IPFS and their identities are recorded on Ethereum-compatible blockchain nodes (Rahman et al., 2023). Across these usages, the common theme is anti-fraud verification, but the underlying objects, methodologies, and threat models differ substantially.

1. Terminological scope and research contexts

The term “FakeChain” does not refer to a single standardized framework across the literature. Instead, it designates at least three separate artifacts. The most prominent is the deepfake benchmark introduced in “FakeChain: Exposing Shallow Cues in Multi-Step Deepfake Detection,” where FakeChain is a benchmark and analysis framework for compositional image forgeries synthesized by chaining Face-Swapping, GAN-based generation, and Diffusion-based generation (Heo et al., 20 Sep 2025). A second usage appears in “FakeScope: Large Multimodal Expert Model for Transparent AI-Generated Image Forensics,” where FakeChain is the paper’s foundational multimodal dataset, consisting of tuples of images, authenticity labels, and long-form forensic reasoning (Li et al., 31 Mar 2025). A third usage is an implementation-oriented reinterpretation of the Verifi-Chain prototype as “FakeChain — a blockchain/IPFS-based system to prevent fake academic certificates,” emphasizing credential verification rather than media forensics (Rahman et al., 2023).

These usages are related primarily by problem class rather than by shared codebase or formal interoperability. The two forensic usages address AI-generated or manipulated imagery, but they target different levels of supervision: one is a benchmark for detector stress-testing under chained manipulations, and the other is a knowledge source for multimodal reasoning and instruction tuning. The credential-verification usage instead addresses forged academic documents through decentralized storage and immutable identity records. This suggests that “FakeChain” functions less as a single research program than as a recurring label for anti-fraud pipelines in distinct domains.

2. FakeChain as a benchmark for multi-step deepfake detection

In the 2025 benchmark paper, FakeChain is defined as “a large-scale benchmark comprising 1-, 2-, and 3-Step forgeries synthesized using five state-of-the-art representative generators,” with the explicit goal of exposing the shallow cues used by current detectors when confronted with hybrid or compositional manipulations (Heo et al., 20 Sep 2025). The benchmark formalizes multi-step manipulation as a composition of generators,

$x_k = f_k(f_{k-1}(\cdots f_1(x_0)\cdots)),$

where the “final manipulation type” is $f_K$, the last generator applied in the chain. Generator tokens are given explicitly as FF (FaceFusion), SG3 (StyleGAN3), SS (StyleSwin), SD3 (Stable Diffusion 3), and SDXL (Stable Diffusion XL), and sequential pipelines are represented with underscore-delimited tokens such as FF_SG3, SG3_SDXL, and FF_SG3_SDXL.

The benchmark is designed around the claim that real-world synthetic content is often repeatedly reused, refined, or repurposed across heterogeneous tools rather than created in a single step. Its central empirical result is that detector performance is dominated by the final manipulation type rather than by cumulative manipulation history. When the last-stage generator at test time differs from the training distribution, F1-score can drop by as much as 58.83%, which the paper interprets as evidence that detectors key on last-stage artifacts rather than on persistent traces from earlier stages (Heo et al., 20 Sep 2025).

The benchmark construction is explicitly enumerated. It contains five single-generator settings, 20 pairwise two-step combinations, and 24 heterogeneous three-stage chains, yielding 49 manipulation combinations in total. For each combination, the dataset contains 25,000 images, for a total of 1,225,000 images. The split is fixed at 70% training, 20% validation, and 10% test per combination. Real faces used for face-swapping are drawn from FFHQ-1024, while diffusion prompts are sampled from the Cartesian product of listed age, ethnicity, and gender attribute sets.

3. Experimental design, analyses, and empirical findings in the benchmark

The evaluation protocol in FakeChain uses three detector families: Xception, F3Net, and MAT. Six training configurations are defined by training on 1-Step and 2-Step variants for each generator family—FS, GAN, and Diff—and detectors are then evaluated on nine categories grouped by final manipulation type and depth: 1-Step FS/GAN/Diff, 2-Step FS/GAN/Diff, and 3-Step FS/GAN/Diff. This yields 54 unique train-test pairs. Metrics are Accuracy, F1-score, and AUC for binary classification, with standard definitions of Precision, Recall, and F1-score given explicitly in the paper (Heo et al., 20 Sep 2025).

Several representative quantitative findings are reported. Xception trained on 1-Step FaceSwap achieves AUC 99.92 and F1 99.35 on 1-Step FS, but only AUC 49.65 and F1 0.44 on 1-Step GAN, and AUC 49.25 and F1 0.30 on 3-Step GAN. The same model yields AUC 59.49 and F1 21.98 on 1-Step Diff, and AUC 61.56 and F1 5.33 on 3-Step Diff. Conversely, Xception trained on 1-Step GAN performs strongly on GAN-final categories—AUC 99.95 and F1 98.99 on 1-Step GAN; AUC 99.95 and F1 99.02 on 2-Step GAN; AUC 99.97 and F1 99.20 on 3-Step GAN—but poorly on 1-Step FS, where AUC is 59.77 and F1 is 8.11. Xception trained on 1-Step Diff achieves AUC 100.00 and F1 100.00 on 1-Step Diff but drops to AUC 49.39 and F1 0.04 on 1-Step GAN. F3Net trained on 2-Step GAN illustrates depth overfitting: it reaches AUC 94.75 and F1 94.87 on 2-Step GAN but falls to AUC 59.64 and F1 39.64 on 1-Step GAN (Heo et al., 20 Sep 2025).

The paper supplements classifier evaluation with feature-space and spectral analyses. t-SNE is applied to penultimate-layer embeddings to visualize separability across manipulation depths and final generators. Spectral analysis converts images to grayscale, subtracts a median-blurred version, computes 2D FFT magnitude spectra, and log-scales them for interpretability. The reported finding is that when GANs or Diffusion models are applied last, their spectral signatures overwrite those from earlier stages, whereas FaceFusion applied last partially preserves residual frequency patterns from previous steps. Mutual Information values quantify erosion of early-stage information; for example, Final FF yields MI(1↔2) = 0.0143 and MI(1↔3) = 0.0121, while Final SD3 yields 0.0133 → 0.0118. These trends are presented as corroborating the overwriting of earlier signals in deeper chains (Heo et al., 20 Sep 2025).

Compression robustness is also evaluated under RAW, JPEG 75, and JPEG 75×2. Using the best-performing training combinations—1-Step FS, 1-Step GAN, and 2-Step Diff—Xception achieves at least 97% accuracy across all depths and final types under RAW, and remains strong under JPEG 75, with 92.18% accuracy on 1-Step GAN. MAT degrades sharply to around 56.42% across conditions, while JPEG 75×2 yields negligible additional degradation compared with single compression in the examples given (Heo et al., 20 Sep 2025).

4. FakeChain as a multimodal forensic reasoning dataset

Within the FakeScope framework, FakeChain denotes a different object: a multimodal dataset intended to teach large multimodal models how to reason about authenticity by explicitly linking “visual trace evidence” to a justified conclusion (Li et al., 31 Mar 2025). The core tuple format is

$\{I, auth, rsn\},$

where $I$ is the image, $auth$ is the authenticity label, and $rsn$ is a structured forensic explanation. The stated purpose is not only to classify images as fake or real, but to encode a human-like cause-to-effect flow from observed abnormalities to an authenticity judgment.

The dataset contains 47,594 tuples, balanced between 23,797 AI-generated images and 23,797 real images. The AI-generated side spans 17 generators: ProGAN, StyleGAN, BigGAN, CogView2, ADM, IF, FuseDream, VQDM, Glide, Stable Diffusion, SDXL, FLUX.1, Adobe Firefly, DALL·E 2, DALL·E 3, Wukong, and Midjourney. The real images are drawn from DIV2K, RAISE, and ImageNet. The taxonomy used to steer reasoning contains 16 evidence categories: texture, edge, clarity, light/shadow, anatomy, layout, symmetry, reflection, perspective, physics, shape, theme, content deficiency, distortion, unrealistic, and overall hue (Li et al., 31 Mar 2025).

The construction methodology is the Anthropomorphic Chain-of-Thought Inference (ACoTI) strategy, organized as Steer, Demonstrate, Enlighten. Thirty trained annotators participate; each image is annotated by two experts, and only the intersection of selected evidence categories is retained. In Step1-Steer, experts judge authenticity and select supporting evidence categories from the standardized taxonomy. In Step2-Demonstrate, 50 fake and 50 real images are fully annotated as exemplar sets $\{\chi^+,\chi^-\}$, with the constraint that every category in the taxonomy appears in at least three exemplars. In Step3-Enlighten, an AI expert generates structured reasoning drafts under few-shot chain-of-thought prompting with structured context including the selected categories $\mathcal{F}_i$, adaptive exemplar subset $\chi_i$, and the correct authenticity label placed in context. Exemplar selection is optimized through the minimum set covering formulation reported in the paper:

$$\begin{aligned} &\chi_{i} = \arg \min_{\chi' \subseteq \{\chi^+ \cup \chi^-\}} |\chi'|, \ &\text{s.t.} \quad L(\chi') = \mathcal{F}_i, \ \chi' \cap \chi^+ \neq \emptyset, \ \chi' \cap \chi^- \neq \emptyset . \end{aligned}$$

The quality-control results emphasize the effect of this human–machine construction process. In a two-alternative forced choice study with six annotators over 100 sampled items, 99.17% of preferences favored ACoTI-generated outputs over non-ACoTI outputs. The paper also reports ablations isolating FakeChain’s contribution to downstream model performance: compared with a baseline without FakeChain or FakeInstruct, “Only FakeChain” improves Detect from 58.22% to 83.27%, Interpret from 0.423 to 0.593, and Reason from 0.296 to 0.576, while Analyze shifts from 0.368 to 0.350 (Li et al., 31 Mar 2025). These figures are used to argue that FakeChain alone substantially improves detecting and reasoning, even before the broader FakeInstruct instruction corpus is added.

5. Role of FakeChain in FakeScope and transparent forensic modeling

FakeChain in FakeScope is not primarily a benchmark for detector robustness; it is a knowledge source for instruction tuning and transparent multimodal reasoning. The paper states that FakeChain’s reasoning texts are transformed into diverse instructional data in VQA and MCQ formats using Mistral and GPT-4o, with deliberate removal of textual contamination to ensure that answers require visual grounding. The resulting FakeInstruct corpus contains approximately 2,000K instructions: approximately 47K absolute authenticity judgments, approximately 95K holistic reasoning items, approximately 715K fine-grained trace-evidence items, and approximately 1,190K extensional knowledge and discussion items (Li et al., 31 Mar 2025).

FakeScope is trained using supervised visual instruction tuning on FakeInstruct, with LLaVA-v1.5 (Vicuna-7B) and mPLUG-Owl2 (LLaMA2-7B) as backbones; the vision encoder is frozen, while the modality projector and LLM are fine-tuned. The paper further introduces a token-based probability estimation strategy that maps anchor tokens “fake” and “real” to logits at the authenticity-token position and computes probabilities through softmax over sums of logits:

$f_K$0

This enables zero-shot quantitative scoring despite training only on qualitative hard labels.

The broader significance of this FakeChain usage lies in its emphasis on interpretability. Prior forensic datasets are described as focusing mainly on binary classification or short template-based explanations, whereas FakeChain is presented as a dataset explicitly organized around “visual trace evidence → authenticity” reasoning. A plausible implication is that this formulation reframes forensic supervision from label prediction toward explanation-conditioned decision support, although the paper also notes uncertainty stemming from reliance on imperfect AI experts for distillation (Li et al., 31 Mar 2025).

6. FakeChain as a blockchain/IPFS credential-verification system

A separate usage of “FakeChain” appears as a technical brief for a blockchain/IPFS-based system to prevent fake academic certificates, aligned with the Verifi-Chain prototype (Rahman et al., 2023). Here the problem setting is credential fraud in Southeast Asia, where submitting fake certificates is described as a common problem that disadvantages qualified candidates, wastes employer time and resources, and erodes trust in academic credentialing. The proposed approach is to design and implement a decentralized certificate verification model leveraging blockchain and IPFS so that certificates become tamper-proof and non-repudiable while storage costs are reduced by keeping large files off-chain.

The operational pipeline is explicitly given. Applicants upload academic credentials through a web interface, and the admin is notified of a verification request. While verification is pending, certificates are temporarily stored in a local database. The admin then contacts issuing educational organizations to confirm authenticity. Once verified, certificates are uploaded to IPFS, which chunks the file, computes a cryptographic content address, and returns a CID. The paper states that “the hash key is encrypted before it is stored in the Blockchain nodes.” After upload, the issuer approves generation charges in MetaMask, and the CID or hash cannot be modified once recorded on-chain. Companies then search for the applicant by the provided hash/CID, request access to view credentials, and after applicant approval, view the certificates (Rahman et al., 2023).

The implementation stack explicitly includes Ethereum smart contracts, Ganache, MetaMask, Truffle, Web3, React, Node.js, JWT, and IPFS. Off-chain storage consists of the certificate file itself in IPFS, while the on-chain component stores the certificate’s unique identity and minimal metadata for verification and access control. A consistent minimal schema proposed in the technical brief includes cid, cidDigest, issuer, subject, issuedAt, status, accessTier, and optionally metadataURI. A certificate hash is described by

$f_K$1

where $f_K$2 is aligned with SHA-256 and $f_K$3 is the canonicalized certificate file bytes as stored in IPFS. The brief additionally presents issuer signatures as an advisable strengthening of non-repudiation:

$f_K$4

with verification

$f_K$5

These digital signatures are described as generally advisable rather than explicitly implemented in the paper (Rahman et al., 2023).

The system’s stated benefits are efficiency, security, and cost-effectiveness. IPFS serves as a “middleman” to avoid the expense of storing massive data directly on-chain, while blockchain immutability and IPFS content addressing make tampering detectable. The brief also presents a standard birthday-bound expression for collision probability,

$f_K$6

for a $f_K$7-bit hash, noting that this is standard cryptographic practice rather than an explicit claim of the original paper (Rahman et al., 2023).

7. Comparative interpretation and research significance

The three usages of FakeChain differ in object type, data model, and validation regime. The deepfake benchmark FakeChain is a large-scale stress test for distribution shift across sequential generators, with explicit quantitative detector evaluations, spectral analysis, t-SNE visualization, and mutual-information analysis (Heo et al., 20 Sep 2025). The FakeChain dataset inside FakeScope is a supervision corpus for explainable multimodal forensics, centered on structured reasoning and human–machine annotation workflows rather than train/test robustness on chained synthesis pipelines (Li et al., 31 Mar 2025). The blockchain/IPFS FakeChain is an application architecture for academic credential verification, where the primary unit is a certificate whose content is stored in IPFS and referenced immutably from blockchain nodes (Rahman et al., 2023).

A concise comparison is useful.

Usage of “FakeChain”	Domain	Primary artifact
Deepfake benchmark	Image forensics	Benchmark and analysis framework for 1-, 2-, and 3-Step forgeries
FakeScope dataset	Multimodal forensics	Dataset of tuples $f_K$8 for authenticity reasoning
Credential system	Document verification	Blockchain/IPFS-based certificate verification workflow

Methodologically, the deepfake benchmark argues that current detectors rely on shallow last-stage fingerprints, as shown by cross-distribution F1 degradation of up to 58.83% and by spectral overwriting effects when GAN or Diffusion stages appear last (Heo et al., 20 Sep 2025). The FakeScope dataset argues that authentic forensic competence requires causal, evidence-grounded explanation rather than only binary prediction, and it reports strong gains from reasoning-centric supervision (Li et al., 31 Mar 2025). The credential-verification system argues that decentralized storage and immutable identity records can reduce forgery risk and manual verification overhead in academic documents (Rahman et al., 2023).

Taken together, these works show that “FakeChain” has become a label for chained evidence, chained manipulation, or chained provenance depending on the domain. In image forensics, the chain is the manipulation history or the reasoning chain; in credential verification, the chain is the blockchain-backed provenance trail. This suggests that the name has converged semantically around verifiable sequences of transformation, evidence, or custody, even though the associated technical stacks are otherwise unrelated.