Image forensics

Image forensics encompasses computational methods to detect image manipulation, tampering, and forgery. This is relevant to misinformation detection because fake news and satire creators often splice, alter, or recontextualize images to support misleading claims or create absurd humor.

Techniques

Error Level Analysis (ELA): Exploits lossy JPEG compression to identify image tampering. The original JPEG image is resaved at a known compression rate, and pixel-by-pixel differences are compared. Regions with low error indicate that the image was already compressed (original), while high error regions suggest newly added or spliced content. Limitations: ineffective on images that have been resaved multiple times (common for internet images) or when the compression level is uniform; cannot identify which manipulation technique was used.

CNN-based detection: Deep learning approaches (e.g., fine-tuned on tampering datasets like CASIA 2.0) learn to classify images as authentic or forged. Often trained on ELA preprocessed inputs.

Splicing detection: Specialized methods to detect image splicing (combining parts from multiple images), a common technique in satirical news.

Applications to misinformation

• Satire detection: Li et al. (2020) applied ELA+CNN to detect manipulated images in satirical news articles as a proxy for satire, though the approach underperformed (44% accuracy), suggesting image tampering alone is insufficient.
• Visual fake news: Detecting doctored images paired with false claims.
• Deep fakes and synthetic media: Advanced forgery detection for synthetic or AI-generated images.

Key papers

• Detecting GAN-generated Imagery using Color Cues — Detects GAN-generated imagery via architectural analysis of generator color formation and normalization, identifying saturation suppression as a distinguishing forensic cue.
• Two-Stream Neural Networks for Tampered Face Detection — two-stream network combining visual artifact detection with steganalysis features for face tampering detection; introduces SwapMe/FaceSwap dataset.
• A Multi-Modal Method for Satire Detection using Textual and Visual Cues: Applies ELA+CNN for satire detection; discusses why image forensics alone fails to reliably identify satire.
• Krawetz (2007): foundational work on Error Level Analysis.
• Dong et al. (2013): CASIA 2.0 image tampering detection dataset.

Connections

• Multimodal fake news detection — image forensics can be combined with text analysis for joint detection.
• Visual misinformation — detecting manipulated images is core to visual fact-checking.
• Deepfakes — advanced synthetic media detection.

Notes

While image forensics is valuable for detecting obvious forgery, it is not a silver bullet for misinformation. Many authentic news images are edited for minor adjustments (color correction, background blur) by reputable outlets like Reuters and AP; conversely, satirical articles may use unmanipulated images from existing sources (e.g., the Joe Rogan podcast example in Li et al. 2020). For reliable detection, joint multi-modal and contextual analysis is necessary.