XAI-Deepfakes¶
Publisher: github.com/IDT-ITI/XAI-Deepfakes | Type: Detector | Cost: Free
TL;DR
A free, open-source academic toolkit from CERTH IDT-ITI that runs Explainable AI methods (GradCAM, SHAP, LIME) on an EfficientNet deepfake classifier to produce visual heatmaps showing which pixels drove the verdict. The primary 1C.3 entry for evidence-grade reporting where a probability score is not enough.
What it does¶
XAI-Deepfakes wraps three Explainable AI methods around an EfficientNet-based deepfake classifier. GradCAM produces gradient- weighted class-activation heatmaps that visualise which regions of the image most contributed to the deepfake verdict; SHAP produces Shapley-value attributions that quantify each pixel's marginal contribution to the classification; LIME produces locally interpretable model-agnostic explanations that approximate the classifier's behaviour on perturbed inputs. The output is a combination of the underlying classifier's verdict plus the three explanation visualisations, which together form an evidence artefact a fact-checker can publish alongside a detection finding and a defamation-defence file can attach to legal pleadings.
The toolkit's institutional value sits in defamation-defence contexts. Where Sensity and Reality Defender produce probabilistic scores plus vendor-defined "explainable indicators" inside a black-box pipeline, XAI-Deepfakes produces the academic state-of- the-art explainability output with reproducible methods. For a court or platform-appeal process that asks "what made the system decide this was AI?", XAI-Deepfakes can answer the question with visualisations that are scientifically grounded rather than vendor-defined.
When to use it¶
- A defamation defence file requires evidence-grade detection output: not just a probability score but a visual map of which pixels drove the verdict, attached to the published debunk as legal-defensibility material.
- A research-grade fact-check needs to show that a Sensity or Reality Defender verdict is corroborated by an academic classifier with explainability output, not only by a black-box vendor pipeline.
- A regional academic partner is producing a forensic report and the methodology must be reproducible by another lab from the published code.
- A graduate-school or partner-university context teaches detector explainability and the GradCAM / SHAP / LIME triad is the methodologically appropriate teaching environment.
Independent accuracy¶
Vendor claim vs independent assessment
Vendor positioning: XAI-Deepfakes is positioned as a methodology toolkit for explainable deepfake forensics rather than as a headline-accuracy classifier. The underlying EfficientNet classifier is a standard academic baseline; the three Explainable AI methods (GradCAM, SHAP, LIME) are canonical in the explainability literature. The maintainer does not market a single headline accuracy figure.
No independent SEA-specific benchmark identified as of May 2026. XAI-Deepfakes is academic tooling and has not been independently audited for SEA-specific performance. The underlying EfficientNet classifier inherits the Deepfake-Eval-2024 benchmark cross- category gap: no Asian-face or SEA-language breakdown is publicly available for image / video detection methodology of this generation. The toolkit's editorial framing is that XAI-Deepfakes' value sits in the explainability output rather than in the underlying classifier's headline accuracy; the classifier should be treated as one weak signal class per Anchor 1, with the explainability output supporting the interpretation work that makes the signal usable in an evidence-grade report.
Limitations¶
Limitations
- Requires Python proficiency, an Ubuntu environment, and a GPU for practical performance.
- Academic tooling, not production. Maintenance follows the research lab's grant cycle rather than a vendor SLA, and operationalising it for a working newsroom requires the newsroom's own technical staff.
- The underlying EfficientNet classifier inherits the same Asian-face / SEA-language benchmark gap that affects every detector in the toolkit (the Deepfake-Eval-2024 cross-category note).
- GradCAM / SHAP / LIME outputs are interpretable but not objective: an explanation map shows what the classifier attended to, not what is true about the image. Treat explanations as evidentiary aids, not as ground truth.
- No Lao-language, Sinhala, or Tamil documentation; the toolkit is English-only academic material.
Privacy and threat model¶
XAI-Deepfakes runs locally on the analyst's GPU machine. The suspect file stays under the analyst's control during analysis; no upload to a vendor cloud is required. This makes the toolkit the right privacy posture for surveillance-environment work in Sri Lanka and Laos (where shipping a suspect file to a vendor cloud would itself constitute a disclosure event) and for defamation-defence work generally, where the suspect file may be sub judice and must remain within the analyst's controlled environment.
The threat model that matters here is downstream: an explanation artefact published alongside a debunk is published research output that adversaries can study. Adversarial deepfake research routinely optimises against published explainability methods to produce content that GradCAM / SHAP / LIME visualise as "real" even when the underlying classifier scores it as fake. Treat explanation outputs as evidence in the published analysis rather than as a permanent solution to detection.
Country and platform applicability¶
- Indonesia: language-agnostic for visual modules; available in principle to Indonesian academic and newsroom- technical partners with GPU infrastructure.
- Laos: language-agnostic; the local-execution model is privacy-defensible. No documented Lao-specific deployment.
- Malaysia: language-agnostic; available to MIMOS, Mesolitica, or university partners with GPU access.
- Philippines: language-agnostic; available to UP / DLSU / Ateneo and to Rappler's technical layer; particularly relevant for defamation-defence work tied to Rappler / #FactsFirstPH published debunks.
- Sri Lanka: language-agnostic for visual modules; LIRNEasia and Watchdog technical capacity can run XAI-Deepfakes inside their research environment; explainability output is particularly valuable under Sri Lanka's OSA framework where evidentiary defence matters.
- Thailand: language-agnostic; available to Thai university partners; relevant under the Thai lèse-majesté and CCA evidentiary regime where explainable detection output reduces legal exposure.
Platform applicability: not platform-specific. Operates on any image or video file the analyst can supply.
How to access¶
Free download from the IDT-ITI XAI-Deepfakes GitHub repository. Clone the repo, install Python dependencies, configure CUDA for GPU support per the README. No account creation, no beta gate. The toolkit is open-source academic code with a permissive research licence.
Cost (current as of 2026-05)¶
Free in software terms. The structural cost is hardware: a GPU (Ubuntu host), at least 16 GB GPU memory recommended for EfficientNet inference plus SHAP attribution at usable speeds. For institutional partners with university or research-lab GPU access, this is a non-issue; for a frontline newsroom without GPU infrastructure, the path is partner-mediated rather than direct.
Quickstart¶
- Confirm an Ubuntu (or Linux compatible) host with a CUDA- capable GPU and at least 16 GB GPU memory.
- Clone the IDT-ITI XAI-Deepfakes repository from GitHub.
- Install Python dependencies and configure CUDA per the README.
- Place the suspect image or video frames in the input directory.
- Run the GradCAM analysis to produce a class-activation heatmap.
- Run the SHAP analysis to produce per-pixel attribution scores (this is the slowest of the three but the most quantitatively defensible).
- Run the LIME analysis to produce a locally interpretable approximation of the classifier's behaviour on the input.
- Combine the three explanation visualisations alongside the underlying EfficientNet score in a single forensic-report artefact for publication or legal use.
In the toolkit's workflow¶
Cautious-detector pillar tool with explanation output per Architectural Anchor 1. Sits in 1C.3 Institutional explainable AI forensics as primary, ahead of TruFor (alternative, non-detector reliability map) and TRIED Benchmark (institutional reference framework, not a deployable tool).
Standard combinations:
- With Sensity at 1C.2 for institutional-pipeline corroboration: Sensity provides the production-grade verdict, XAI-Deepfakes provides the academic explanation artefact that defamation defence can attach.
- With TruFor at 1C.3 for parallel non-detector reliability-map output on the same suspect image; the two together give per-pixel explanation (XAI-Deepfakes) plus reliability-of-the-reliability-map (TruFor).
- With DeepfakeBench at 1C.2 inside a research- evaluation context where XAI-Deepfakes' EfficientNet baseline is being compared against the harness's 36 detectors.
- With TRIED Benchmark at 1C.3 as the sociotechnical evaluation framework that complements XAI-Deepfakes' technical explanation output.
This tool's verdict is one weak signal class per Anchor 1; never publish a binary claim from it alone. The explanation visualisation adds evidentiary weight to a published claim but does not by itself constitute a non-detector signal under Anchor 2's two-non-detector-signal floor.
Decision-tree references: T6 source-protection routes through XAI-Deepfakes' local-execution model when the suspect file cannot leave the analyst's machine. T5 escalation references the explanation artefact at T5.15 (method note) when the published debunk needs a defamation-defence layer.
Conflict resolution behaviour¶
When XAI-Deepfakes' verdict disagrees with another detector (Sensity, Reality Defender, Hive), XAI-Deepfakes carries weaker headline-accuracy weight than the enterprise platforms because its underlying EfficientNet classifier is an academic baseline rather than a Sensity-grade ensemble. It carries stronger evidentiary weight than the enterprise platforms in defamation-defence contexts because the explanation output is reproducible and methodologically grounded, where vendor "explainable indicators" are vendor-defined and not openly reproducible. When XAI-Deepfakes disagrees with a non-detector signal (provenance manifest, reverse-image hit, archived earlier instance), the non-detector signal wins because Anchor 2 mandates two non-detector signals before any strong public claim. The explanation artefact is most useful when it visualises why the underlying detector and a non-detector signal agree, not when it tries to substitute for the non-detector signal.
Override notes¶
Override notes
-
Cautious-detector with explanation (g1-cautious-detector- with-explanation): XAI-Deepfakes serves the cautious- detector pillar with an explanation-output supplement. The detector verdict is one weak signal class per Anchor 1; the explanation output supports interpretation rather than adding an independent signal.
-
Defamation-defence pointer (defamation-defence-use-case- pointer): the toolkit's published-output use case targets defamation defence and platform-appeal evidentiary work. Country and platform applicability section above names the Sri Lanka OSA, Thai lèse-majesté and CCA, and Philippine COMELEC contexts where explainability output materially reduces legal exposure on a published debunk (regional legal-context research).
Sources¶
- IDT-ITI, Centre for Research and Technology Hellas. XAI-Deepfakes — explainable deepfake detection. GitHub repository (open-source, CUDA/Ubuntu). github.com/IDT-ITI/XAI-Deepfakes.