38 detection modules working together to authenticate digital media. Each module examines a different aspect of a file — from pixel patterns to audio signals — so nothing escapes analysis.
These modules examine the internal structure of media files — how frames are organized, how data is encoded, and whether the technical plumbing of a file tells a consistent story.
Checks whether the frame grouping pattern (I-frames, P-frames, B-frames) is consistent throughout the video.
How it works: Edited videos often break the regular repeating pattern of frame groups. This module maps the entire GOP structure and flags any irregularities.
Verifies that each frame's timestamp is evenly spaced and consistent with the stated frame rate.
How it works: Examines the presentation and decoding timestamps (PTS/DTS) of every frame. Spliced or inserted footage often creates timing jumps or gaps.
Reads basic file metadata to check whether creation dates, software tags, and encoding settings appear consistent.
How it works: Extracts standard metadata fields and looks for signs that the file was created or modified by different tools than expected.
Performs 8 in-depth metadata checks using multiple forensic tools to cross-reference internal file information.
How it works: Compares results from ExifTool and MediaInfo, looking for contradictions between metadata fields that would indicate tampering or re-encoding.
Analyzes bitrate fluctuations, encoding parameters, and motion vectors to find segments that were altered.
How it works: Edited or spliced sections typically show abrupt changes in bitrate, quantization, or motion patterns that differ from the surrounding footage.
Identifies the exact codec, profile, and level used to encode the video and checks for mismatches.
How it works: Compares declared codec information against actual encoding artifacts. Re-encoded or transcoded files often carry telltale codec inconsistencies.
Inspects the raw bitstream for irregularities in how video data is packaged at the lowest level.
How it works: Parses Network Abstraction Layer (NAL) units to detect structural anomalies, out-of-sequence parameters, or signs that raw video data was manually altered.
Measures whether the audio track stays precisely synchronized with the video throughout the file.
How it works: Compares audio and video stream timing and duration. Edited files frequently have subtle sync drift where audio and video no longer line up exactly.
These modules look at individual pixels and their mathematical relationships. Editing an image or video frame leaves subtle traces in how pixels relate to each other — traces that are invisible to the eye but detectable through analysis.
Examines the unique sensor noise pattern that every camera leaves on its recordings, like a digital fingerprint.
How it works: Every camera sensor has tiny manufacturing imperfections that create a consistent noise pattern. Edited regions disrupt this pattern, revealing where changes were made.
Runs Error Level Analysis, JPEG ghost detection, chromatic aberration checks, and CFA pattern analysis on still images.
How it works: Re-saved or composited regions compress differently than original content. This module highlights areas that have been saved more or fewer times than the rest of the image.
Analyzes compression artifacts across the file to detect whether different parts were compressed at different quality levels.
How it works: When edited content is pasted in from another source, it often carries a different compression history. This module detects those mismatched quality levels.
Converts image data into frequency space to reveal hidden periodic patterns that indicate manipulation.
How it works: Uses Fourier transforms to look for unnatural frequency signatures. Resizing, rotation, cloning, and GAN-generated content all leave distinctive frequency fingerprints.
Finds regions within the same image or frame that have been duplicated to cover up or fabricate content.
How it works: Compares small image patches against each other to find near-duplicate regions. Even if rotated or scaled, copied areas share enough features to be detected.
Tracks how objects move between frames and flags motion that is physically impossible or inconsistent.
How it works: Calculates the pixel-level motion field between consecutive frames. Spliced or AI-generated footage often has motion discontinuities that real-world physics would not produce.
Generates visual fingerprints of frames to detect near-duplicate content and track how much a file has been altered.
How it works: Creates compact hash values based on visual content rather than raw data. Two images that look similar will have similar hashes, even if the file data differs.
These modules use artificial intelligence and machine learning to identify AI-generated or AI-manipulated content. They combine local neural network models running on your machine with cloud-based AI consensus from multiple providers.
Runs 6 specialized neural network models locally on your computer to detect deepfakes and AI-generated media.
How it works: An ensemble of ONNX models each analyze frames from different angles — face manipulation, generation artifacts, texture anomalies — and their combined scores provide a robust AI detection result.
Sends sampled frames to 6 independent cloud AI services and compares their analysis for a consensus verdict.
How it works: Multiple leading AI platforms each independently evaluate the media. When multiple providers agree something is manipulated, the confidence level is significantly higher than any single system.
Detects the invisible fingerprints left by Generative Adversarial Networks (GANs) used to create fake images and video.
How it works: GANs leave characteristic spectral patterns, upsampling artifacts, and channel correlations in the content they generate. This module identifies those telltale signatures.
Identifies content created entirely by AI systems, including diffusion models, neural rendering, and other synthesis methods.
How it works: Analyzes noise distributions, spectral patterns, optical flow, and inter-pixel correlations that differ between real-world captures and synthetically generated content.
Uses results from all other modules to perform a focused, intelligent re-analysis of the most suspicious areas.
How it works: After all modules complete their initial analysis, this module identifies areas where multiple tests flagged concerns and performs deeper cross-referenced examination.
Research-grade deep learning models from the academic forensics community, running locally with GPU acceleration. These modules detect manipulation at a level of precision that traditional algorithms cannot reach.
Uses a deep neural network to extract and analyze the unique noise fingerprint left by a camera's sensor.
How it works: A trained CNN learns to separate camera-specific noise from image content. Manipulated regions lose their original camera fingerprint, making edits visible in the residual noise map.
Pinpoints the exact regions within an image that have been manipulated, down to the pixel level.
How it works: The ManTra-Net architecture detects 385 types of manipulation traces. It produces a heatmap showing exactly where in an image changes were made, regardless of the editing technique used.
A general-purpose convolutional neural network trained to distinguish real media from any type of AI-generated content.
How it works: Trained on a broad dataset of both real and AI-generated media, this model learns universal features that distinguish authentic content from synthetic, regardless of the specific AI tool used.
Uses OpenAI's CLIP vision-language model to detect semantic inconsistencies that indicate AI generation or manipulation.
How it works: CLIP understands both visual content and natural language. It can identify high-level semantic anomalies — such as impossible scenes or inconsistent contexts — that pixel-level analysis might miss.
Analyzes whether visual features remain consistent from frame to frame across the entire video timeline.
How it works: Extracts deep features from consecutive frames and measures how smoothly they change over time. AI-generated or spliced video often has subtle frame-to-frame flickering or inconsistency that real footage does not.
Examines video as a 3D volume (width, height, and time) to detect manipulation patterns that span multiple frames.
How it works: Uses 3D convolutional networks to analyze spatial and temporal dimensions simultaneously. This catches editing artifacts that only appear when you look at how pixels change across both space and time together.
Specialized modules that focus on human faces and biological signals. Deepfakes most commonly target faces, so these modules provide critical detection for the most prevalent type of media manipulation.
Checks for a real human heartbeat signal hidden in subtle skin color changes visible in video of a person's face.
How it works: Real people show tiny, rhythmic color changes in their skin caused by blood flow (remote photoplethysmography). Deepfake faces lack this biological signal because they were never alive.
Measures whether lip movements precisely match the spoken audio, detecting dubbed or AI-generated speech.
How it works: Extracts lip position data from video and phoneme timing from audio, then correlates them. AI-generated lip movements often lag, overshoot, or miss the nuances of natural speech.
Identifies when beauty filters, face-altering apps, or cosmetic AI filters have been applied to change someone's appearance.
How it works: Analyzes skin texture, facial geometry, and edge patterns around facial features. Filters typically smooth skin unnaturally and alter proportions in ways that leave detectable traces.
These modules verify provenance, detect hidden data, and check whether content has been visually altered through overlays, annotations, or watermarking. They answer the question: is this content what it claims to be?
Checks for Content Credentials (C2PA) provenance data and invisible watermarks that prove where content originated.
How it works: Reads C2PA manifests embedded by cameras and software, and detects invisible watermarks applied by content platforms. Verified provenance is strong evidence of authenticity.
Detects watermarks embedded by AI generation tools that indicate content was created by artificial intelligence.
How it works: Many AI image and video generators embed invisible neural watermarks in their output. This module identifies those watermarks, proving the content was machine-generated.
Detects hidden data concealed within media files, such as secret messages or files embedded in images.
How it works: Uses multiple steganalysis methods (RS, SPA, DCT, bitplane, and SPAM analysis) to detect statistical anomalies that indicate data has been secretly embedded within the media.
Finds text overlays, redaction boxes, blur effects, and graphic annotations that have been added on top of original content.
How it works: Analyzes edge patterns, color distributions, and texture changes to distinguish original scene content from overlaid graphics, text, or blur effects applied after capture.
Determines which specific camera or device captured the media by analyzing its unique sensor characteristics.
How it works: Matches sensor noise patterns, lens distortion profiles, and processing signatures against known camera models to verify or identify the source device.
Dedicated modules for analyzing audio tracks. From detecting voice cloning and AI-generated speech to finding splices in recordings, these modules examine the audio dimension of media authenticity.
Comprehensive audio analysis including electrical network frequency (ENF) verification, splice detection, noise floor analysis, and text-to-speech identification.
How it works: Authentic recordings carry a consistent hum from the electrical grid (ENF). This module checks that signal, looks for discontinuities indicating splices, and identifies synthesized speech patterns.
Uses AI to determine whether a voice recording is from a real person or was generated by a voice cloning or text-to-speech system.
How it works: Analyzes spectral characteristics, micro-prosody, and temporal patterns in speech that differ between genuine human voices and AI-synthesized audio, even from the most advanced voice cloning models.
Specialized modules for analyzing surveillance and security camera footage. These handle the unique challenges of CCTV evidence, including proprietary formats, burned-in timestamps, and common tampering techniques.
Reads and verifies on-screen display (OSD) timestamps burned into security camera footage to detect time gaps or alterations.
How it works: Uses OCR to read embedded timestamps frame by frame, then checks for gaps, jumps, or inconsistencies in the time sequence that would indicate footage was deleted or rearranged.
Runs a standardized set of media authentication tests (MAT v6) used by forensic examiners to evaluate video evidence.
How it works: Executes a formal suite of pass/fail tests based on established forensic standards for media authentication, producing results in a format recognized by forensic laboratories and courts.
Detailed technical documentation and accuracy metrics are available to registered users. Download Forensic Media Analyzer to access comprehensive module documentation, configure detection thresholds, and run these modules on your own evidence files.
Download FMA