Our Mission
FakeRadar was created to address a growing problem: AI-generated images and deepfake videos are becoming indistinguishable from real content, yet the tools to detect them have remained locked behind enterprise paywalls or academic research.
Our mission is to make forensic-grade AI content verification accessible to everyone — journalists fact-checking viral images, researchers studying synthetic media, and individuals who simply want to know if what they're looking at is real.
We don't claim to provide definitive answers. We provide forensic signals. The interpretation remains yours.
Our Methodology
FakeRadar applies a multi-signal approach to content verification. No single detector is definitive — AI models make mistakes, ELA can be fooled by re-saving, and metadata can be stripped. That's why we layer multiple independent signals and present them together.
Our analysis pipeline works as follows:
- 1 SHA-256 fingerprinting — The uploaded file is hashed before analysis. If we've seen this exact file before, we return the cached result instantly without re-processing. This eliminates redundant API calls and speeds up repeat queries.
- 2 Multi-engine AI detection — The file is sent simultaneously to Hive AI (primary) and Sightengine (Pro). Each model returns independent confidence scores trained on different datasets and architectures. Agreement between models strengthens confidence; disagreement signals uncertainty.
- 3 Forensic signal analysis (Pro) — ELA, FFT frequency analysis, C2PA credential check, and EXIF metadata extraction run in parallel on our inference server. These signals operate entirely independently of the AI detectors.
- 4 Signal aggregation — Results from all engines are aggregated and presented as a unified report. Each signal is shown with its raw confidence value, not summarized into a single opaque score.
We deliberately avoid a single "real/fake" verdict because no detector achieves 100% accuracy on all content types. Instead, we present each signal with its confidence level so you can make an informed judgment.
Detection Signals Explained
AI Model Detection
Deep learning classifiers trained on millions of real and AI-generated images. These models learn statistical patterns that differ between camera-captured photos and generative model outputs (GAN artifacts, diffusion model fingerprints). Hive AI and Sightengine each use independently trained architectures, reducing the risk of shared blind spots.
Error Level Analysis
ELA re-saves an image at a known compression level and measures the difference between the original and re-saved version. Authentic photos compress uniformly — edited or composited regions compress differently because they've been processed a different number of times. ELA heatmaps make these inconsistencies visible.
Frequency Domain Analysis
Fast Fourier Transform analysis converts an image from pixel space to frequency space. Many AI-generated images contain characteristic frequency artifacts — periodic patterns or unusual spectral distributions — that are invisible to the eye but appear clearly in the frequency domain. This technique is particularly effective against GAN-generated images.
Content Credentials (C2PA)
C2PA (Coalition for Content Provenance and Authenticity) is an open standard for embedding cryptographically signed provenance data into media files. When a camera, phone, or AI platform embeds C2PA credentials, FakeRadar verifies the cryptographic signature and displays the full provenance chain — device model, capture time, software used, and any edits applied.
EXIF Metadata
EXIF metadata records camera settings, GPS coordinates, timestamps, and software used when a photo is taken. AI-generated images typically lack authentic EXIF data, or contain metadata inconsistencies — such as software signatures from image editors applied to a file with a camera model that never existed. We inspect and surface these anomalies.
No single signal is conclusive. EXIF can be stripped; ELA can be confused by heavy compression; AI models have false positive rates. Use FakeRadar results as one input in a broader verification process, not as final proof.
Who It's For
Journalists & Fact-Checkers
Verify viral images and video clips before publication. Generate shareable analysis reports with a permanent link to document your verification process. Our journalist-specific guide covers recommended workflow and limitations.
Researchers
Study AI-generated media patterns with access to raw signal data. Pro tier provides ELA heatmaps, FFT spectra, and Sightengine scores alongside Hive results — suitable for comparative analysis across multiple detectors.
Social Media Moderators
Quickly assess flagged content before escalation. The analysis runs in seconds and generates a timestamped report that can serve as documentation for moderation decisions.
Individuals
Anyone who encounters a suspicious image or video online. Free tier covers basic AI detection for images — no account required for your first analysis.
Technology Stack
FakeRadar is built on Cloudflare's global edge network, with analysis processing distributed to minimize latency worldwide.
The forensic analysis server (ELA, FFT, C2PA, EXIF) runs our own Python implementation, not a third-party service. This gives us full control over the analysis pipeline and ensures no uploaded content is sent to additional third parties beyond what's documented in our Privacy Policy.
Privacy & Zero Retention
FakeRadar does not store your uploaded files. Here's exactly what happens to your content:
- Upload: Your file is received by our Cloudflare Worker, processed in memory, and forwarded to the analysis engines.
- Analysis engines: Hive AI and Sightengine receive your file for classification. Their data handling is governed by their own privacy policies, which we link in ours.
- Forensic server: For ELA and FFT analysis, the file is sent to our private inference server. The original file is discarded immediately after analysis. Only the generated heatmap images are stored temporarily in R2 (up to 90 days for Pro, shorter for free) for display in your report.
- What we keep: A SHA-256 hash of the file content (used to cache results and avoid re-processing identical files), the analysis scores, and metadata like file type and dimensions. Never the original file.
- Shared reports: If you share a result, the report remains accessible at its URL. You can delete shared reports from your dashboard at any time.
For complete details, see our Privacy Policy.
Contact & Background
FakeRadar is built and maintained by Oktay Atalay, a creative designer working in Turkey who has made images and video professionally since 2015, in pharmaceutical communications and international event production. Day to day that means key visuals, campaign material, video editing and 3D work. The practice of making images, in other words, rather than the practice of investigating them.
That background is why this tool exists. When generative models got good enough that a synthetic photograph could pass for a commissioned one, the question stopped being academic for anyone whose job involves deciding whether an image can be used. The available answers were enterprise contracts or research code run from a terminal, and neither is reachable for a person holding one suspicious picture.
I am not a security researcher and do not present myself as one. The detection models come from Hive and Sightengine. The forensic methods, ELA, FFT, C2PA and EXIF inspection, are published standard techniques rather than anything invented here. What is built on this site is the layer that makes them usable together, honest about where each one fails, and reachable without an account. The guides section documents the reasoning behind every signal reported here, including the cases where a signal should not be trusted.
Author profile: Oktay Atalay · LinkedIn
If you've found an analysis error, a bug, or want to discuss a partnership or press inquiry: