不可視AI透かし＆周波数スペクトラム鑑識スキャナー
Google SynthIDのフーリエ同心円リング、生成AI特有の格子アーティファクト、LSBノイズプレーン、ELA圧縮境界、16–24kHz超音波透かしをブラウザ上で直接可視化。
画像または音声ファイルをドロップして解析
対応形式: JPEG, PNG, WebP, AVIF, TIFF, BMP, MP3, WAV, M4A, FLAC, OGG。最大1024×1024 Float32Array Web Worker（メモリ約8MB）。
完全端末内処理: ファイルはサーバーへ一切送信されず、すべてのFFT演算はデバイスのブラウザ上で実行されます。
2D 高速フーリエ変換 (FFT) 設定
Forensic Signal Metrics
Executive Forensic Triage: How to Detect AI Images & Invisible Watermarks
Instant answers synthesized for researchers, forensic analysts, and search engines verifying synthetic media integrity.
Is this image AI generated?
To verify if an image is AI generated, apply a Two-Layer Forensic Audit: First, inspect Layer 1 (Container Provenance) for C2PA JUMBF manifests, IPTC DigitalSourceType tags, and camera EXIF headers. Second, inspect Layer 2 (Signal Mathematics) using a 2D Fast Fourier Transform (2D FFT) to identify periodic lattice crosshairs from neural upscalers, concentric Fourier carrier rings from latent watermarks (SynthID), and flattened Shannon bit-entropy (< 0.95) in the Least Significant Bit (LSB) plane.
What is an invisible AI frequency watermark (SynthID)?
An invisible AI frequency watermark (e.g., Google DeepMind SynthID, Meta Stable Signature) is an imperceptible signal embedded directly into the mathematical coefficients of an image (Discrete Wavelet Transform sub-bands or Fourier coefficients) or audio waveform during synthesis. Unlike EXIF or C2PA metadata stored in file headers, frequency watermarks alter spatial and spectral relationships, surviving screenshots, cropping, lossy JPEG recompression, and color grading. In 2D FFT spectral analysis, they emerge as circular rings or harmonic halos around the DC center.
How can you detect AI images when metadata is stripped?
When metadata is erased by social networks or screenshotting, detect AI images through frequency-domain and steganographic inspection: (1) 2D FFT exposes high-frequency grid artifacts and starburst crosshairs caused by transposed convolutions and pixel-shuffle decoders; (2) LSB Bit-Plane Slicing calculates Shannon entropy on Bit 0—natural camera sensors exhibit chaotic Poisson noise (~0.98–1.00), while AI generators leave flat zero-entropy dead zones; (3) Error Level Analysis (ELA) exposes localized recompression error disparities caused by AI inpainting or Generative Fill.
Container Provenance (Layer 1) vs Signal Forensics (Layer 2)
Why inspecting metadata alone is no longer enough to detect modern generative AI images and synthetic audio.
| Inspection Dimension | FreeImGen AI Signal Scanner | Layer 1 Metadata (C2PA/EXIF) | Traditional ELA Tools |
|---|---|---|---|
| Google SynthID Detection | 2D FFT Frequency Rings | Invisible to headers | Cannot detect |
| Survives Screenshot / Re-save | Embedded in pixel math | 100% stripped | ~ Degraded by re-save |
| Diffusion Upscaler Artifacts | Crosshairs & harmonic grids | No signal | ~ Inconclusive |
| Suno / Udio Ultrasonic Watermarks | 16–24 kHz STFT Waterfall | Missing in MP3s | Images only |
| Zero-Upload Privacy | Pure Client-Side TS | Local parser | Most upload to cloud |
AI Provenance & Watermark Ecosystem Tools
Seamlessly transition between metadata container inspection, C2PA manifest removal, and signal-level frequency forensics.
AI Metadata Checker
Inspect C2PA JUMBF containers, IPTC DigitalSourceType, and camera EXIF headers before running frequency analysis.
Remove Content Credentials
Sanitize C2PA cryptographic manifests and Adobe Content Credentials without re-encoding or degrading pixel quality.
Instagram AI Info Label Remover
Strip false-flag metadata tags triggering Meta automated "AI Info" badges on genuine photos edited in Lightroom.
Why Instagram Says AI Info on Real Photos
Read the comprehensive investigation explaining how phone camera AI and subtle retouching trigger automated platform warnings.
技術解説: 周波数領域ステガノグラフィの数学
Cooley-Tukey 2D 高速フーリエ変換 (FFT)
画素空間の輝度を二次元周波数成分に変換し、中心を直流 (DC) とするfftshiftを行うことで、不可視な変調信号を視覚的ヒートマップとして顕在化させます。
不可視AI透かし＆周波数鑑識に関するよくある質問
C2PAメタデータとSynthID等の信号透かしの違いは何ですか？
C2PAやEXIFなどのコンテナメタデータはファイルヘッダーに記述されており、スクリーンショット等で容易に消失します。一方、Google SynthIDやAudioSealなどの信号透かしは、画素や波形の周波数係数（フーリエ変換/ウェーブレット変換係数）そのものに直接埋め込まれているため、再圧縮やトリミングにも耐性があります。
アップロードした画像や音声はサーバーに送信されますか？
いいえ、一切送信されません。すべての演算は純粋なTypeScriptとWeb Workerにより、ユーザーのブラウザ内部でのみ完結します。