FAKECHECK analyzes images and video with specialized detection models, then explains the visual and temporal signals behind each score.
Deepfakes aren’t all created the same way.
Deepfake technology has rapidly expanded beyond face swap into generative AI images and video, and adversarial manipulation designed to evade detection models. Different manipulation methods leave different traces in an image and require different features to detect. Different manipulation, different detection.
Swapping an existing face
Replaces or blends a real person's face with someone else's.
Generating a person who doesn't exist
Generative AI creates entirely new faces and scenes from scratch.
Manipulated to fool detection models
Inserts noise and distortions nearly invisible to the naked eye to evade the judgment of general detection models.
In deepfake detection, the data makes the difference.
Deepfake detection models are heavily influenced by the domain and facial characteristics of their training data. FAKECHECK has advanced its detection models by reinforcing them with data from real target environments and reflecting Korean facial characteristics across a range of age groups.
A domain included in training
Shows a relatively high detection rate for domains included in training.
A domain not seen in training
Detection rates can drop for new domains not included in training.
Training across diverse domains
The more broadly a model is trained across diverse domains, the better it responds to new ones.
Korean facial characteristics, built into real training data.
Beyond public datasets, we use Korean facial images collected with explicit consent to improve detection performance for specific age groups and Korean individuals.
Specialized models analyze different types of manipulation.
Instead of judging every deepfake with a single model, FAKECHECK separates and combines purpose-specific detection models matched to the characteristics of each different generation and manipulation method.
Detects traces of a swapped face
Analyzes the facial region and key features to detect signs of FaceSwap-based manipulation.
Detects AI-generated people
Analyzes the characteristics of facial images created by a range of generative AI tools, including ChatGPT-family image generation features.
Detection tuned to Korean facial characteristics
Reflects Korean faces and data across a range of age groups in training to improve detection in real domestic use environments.
Analyzes even manipulation designed to fool detectors
Analyzes the subtle noise and adversarial distortion patterns designed to keep general detection models from flagging content as fake.
Detect deepfakes designed to evade detection.
Deepfake attacks no longer stop at making convincing content. Attacks are now emerging that insert noise and distortion, nearly imperceptible to a human, into an image specifically to evade the judgment of the detection model itself.
The original manipulated image
A deepfake image created using face swap or generative AI.
Subtle distortion added to evade detection
Noise and distortion, nearly imperceptible to the naked eye, are added to the image, designed to evade the detection model's judgment.
A general detector can misjudge it
A general detection model that isn't built to counter this kind of manipulation may misjudge the image as real.
Extending analysis to adversarial patterns
FAKECHECK is extending its analysis scope to include adversarial images specifically crafted to evade deepfake detection.
More than a fake-or-real result. See the evidence behind it.
In deepfake detection, the reasoning behind a verdict matters as much as the verdict itself. FAKECHECK visualizes the detection probability and the regions the model focused on, and provides the analysis results as a natural-language explanation people can understand.
Deepfake probability
Provides each detection model's analysis result and Fake Probability score.
Pixel & facial-region visualization
Uses an XAI-based heatmap to visually show the regions the model focused on.
Facial-feature-based analysis
Connects the results to key facial features and anomaly patterns around the eyes, nose, mouth, and more.
Natural-language explanation of the reasoning
Synthesizes the detection result and visual evidence to explain, in natural language, why an image was judged fake.
From a single image to temporal patterns across video.
From a single image to the full reasoning
Detects the facial region and analyzes it for FaceSwap or generative AI origin, providing the deepfake probability, per-model results, heatmap, and natural-language explanation together.
Analyzing the full timeline of a video
Extracts frames from a video and analyzes the facial region to identify the deepfake probability over time, suspicious segments, and representative frames. The supported formats and analysis scope of the current commercial service should be confirmed against the latest product policy.
From upload to evidence, in one analysis workflow.
Image/video input
The user uploads the image or video to analyze.
Face/frame preprocessing
Frames are extracted from video, and the analyzable facial region is detected in the image.
AI detection by manipulation type
Purpose-specific models — for FaceSwap, generative AI, evasion, and more — each analyze their own type of manipulation trace.
Evidence analysis
Synthesizes the detection probability, XAI-based activation regions, and facial features.
Deliver results and reasoning
Visualizes the detection result and delivers it with an LLM-based natural-language explanation the user can understand.
Deploy Fakecheck wherever content authenticity matters.
Identity verification & KYC
Analyzes facial images and ID photos submitted during sign-up and verification for signs of manipulation, strengthening identity verification.
Media & fact-checking
Analyzes person-related images and videos circulating in news and social media for deepfake probability, supporting content verification and fact-checking.
Countering brand & executive impersonation
Analyzes fake content and impersonation images that misuse the faces of CEOs, employees, and public figures, supporting brand trust and damage response.
Digital crime & public safety
Supports rapid authenticity checks in situations of social harm involving deepfakes — impersonation, disinformation, digital sex crimes, and more.
Analyze directly— or integrate Fakecheck into your service.
Analyze in the Web App
Users can upload images or videos directly and check the analysis results.
Integrate via API
Provides a structure for sending files via REST API and receiving deepfake analysis results in JSON format.
Integrates with a range of business systems
Can connect to a range of workflows — KYC, content moderation, media verification, fraud detection, internal investigation, and more.
Detection performance measured under transparent test conditions.
Results by manipulation type
Detection performance is measured separately by manipulation type — FaceSwap, generative AI, adversarial manipulation, and more.
Transparent test conditions
Verifies performance while disclosing test conditions such as dataset composition, sample count, and measurement date.
Validation on Korean Face Data
Continuously validates performance based on Korean facial data across a range of age groups.
Continuous performance improvement
Continues to advance detection models and performance validation in response to new generation tools and manipulation methods.
Frequently asked questions about Fakecheck.
Why does Fakecheck use multiple detection models?
FaceSwap, generative AI images, and detection-evading manipulation are created differently and leave different traces in an image. Fakecheck separates purpose-specific models matched to each manipulation method to analyze the characteristics of each type.
Why does Korean-specialized data matter?
Deepfake detection models are influenced by the facial characteristics and domain of their training data. Fakecheck has been developed to reinforce training with Korean faces and data across a range of age groups, improving detection performance in real domestic environments.
Can Fakecheck detect fully AI-generated images?
Yes. In addition to FaceSwap, which swaps an existing face, Fakecheck analyzes newly generated person images from generative AI with a separate, dedicated model.
Does Fakecheck provide more than a fake-or-real result?
No — Fakecheck also visualizes, via XAI, the image regions the model weighted most heavily alongside the detection result, and provides the result as a natural-language explanation people can understand.
Can Fakecheck analyze video as well as images?
Fakecheck has developed technology that extracts frames from video and analyzes the facial region to identify deepfake probability over time and suspicious segments. The supported formats and analysis scope of the current commercial service should be confirmed against the latest product policy.
Can Fakecheck integrate with our existing service via API?
Existing technical materials describe an integration structure that sends files to be analyzed via REST API and returns results in JSON format. At the time of actual adoption, please confirm the current API delivery method and terms of use.
Can I trust the detection result 100%?
Deepfake detection is a probabilistic result from an AI model and can be affected by various conditions — image quality, compression, manipulation method, new generative models, and more. Fakecheck provides multiple expert models and the reasoning behind each result together, so users can make a more comprehensive judgment.
Go beyond a fake-or-real result. See the evidence behind it.
Evaluate Fakecheck with your own images, video, and operating requirements.
How does this product behave in your environment?
Whether you're exploring, evaluating, or rolling out, you connect directly with a SANDS Lab solutions engineer. Clear every question before contract — that's the point.
Product evaluation & PoC
Real-data PoCs, technical deep-dive sessions, and custom integration scoping. Everything you'd need to validate technical fit before the paperwork starts.
Technical collaboration & licensing
If you want to use the product in an academic benchmark or co-authored paper, we support research licenses and the underlying datasets. Co-authorship is on the table.