Media forensics glossary

A deliberate manipulation of input data designed to fool a machine learning model into producing incorrect outputs. In media forensics, adversarial attacks target detection systems by adding imperceptible noise or perturbations to synthetic content so that classifiers misidentify it as authentic. Common techniques include FGSM (Fast Gradient Sign Method), PGD (Projected Gradient Descent), and diffusion-based adversarial perturbations. See also: GAN fingerprint, deepfake.

The documented history of how an AI system produced a particular piece of content, including the model used, training data origins, generation parameters, and any post-processing applied. AI provenance goes beyond simple labeling by creating a verifiable chain of information about the generation process. Related to but distinct from content provenance, which covers all content regardless of origin. Full guide.

The scientific examination of audio recordings to determine authenticity, detect manipulation, and extract identifying information. Techniques include ENF analysis, spectrogram inspection, MFCC extraction, formant analysis, and compression artifact detection. Modern audio forensics increasingly focuses on identifying AI-generated speech and voice clones. Full guide.

A component in neural networks that allows models to focus on specific parts of the input when generating output. In transformer-based generators, attention mechanisms determine which input tokens influence each output token. Forensically relevant because attention patterns leave statistical signatures in generated text and can help distinguish between different model architectures.

The process of verifying that a piece of digital content is genuine and has not been tampered with since its creation. Content authentication can rely on metadata analysis, cryptographic signatures (as in C2PA), forensic examination, or a combination of these approaches. Differs from identity authentication, which verifies the person rather than the content.

A neural network architecture that learns to compress input into a compact representation (encoding) and then reconstruct it (decoding). Autoencoders are used in early face-swap deepfakes, where two autoencoders share an encoder but use separate decoders for each face. The encoding-decoding process introduces artifacts detectable by forensic analysis, particularly around facial boundaries and fine details.

A category of deepfake detection techniques that look for the presence or absence of physiological signals in video content. These include natural blinking patterns, subtle skin color changes caused by blood flow (remote photoplethysmography), eye gaze consistency, and breathing-related micro-movements. Current generation deepfakes often fail to reproduce these signals accurately, making them useful forensic markers. See: deepfake detection.

A forensic detection approach that works without any reference to the original, unmodified content. Blind detectors must assess authenticity based solely on the content as presented, using statistical analysis, learned features, and artifact detection. Most real-world forensic scenarios require blind detection because the original version is unavailable for comparison.

The use of distributed ledger technology to create tamper-evident records of content creation and modification. Each transaction (creation, edit, transfer) is recorded as a block linked to previous entries via cryptographic hashes. While blockchain offers strong immutability guarantees, challenges include storage costs, throughput limitations, and the fundamental issue that a blockchain can only verify that a record was made, not that the record's contents are truthful.

A statistical measure of how much sentence length and complexity vary within a piece of writing. Human writing typically shows high burstiness, mixing short punchy sentences with longer complex ones, while AI-generated text tends toward more uniform sentence structures. Burstiness is one of several statistical features used in stylometry and AI text detection alongside perplexity and type-token ratio. Related guide.

A technical standard developed jointly by Adobe, Microsoft, Intel, and other organizations for attaching tamper-evident provenance metadata to digital content. C2PA specifies how to create, embed, and verify Content Credentials using cryptographic signatures stored in JUMBF containers. The standard is maintained by the C2PA Working Group under the Linux Foundation's Joint Development Foundation. Limitations include voluntary adoption and vulnerability to metadata stripping. Full guide.

A trusted organization that issues digital certificates used to verify the identity of signers in cryptographic systems. In C2PA, certificate authorities anchor the trust chain that validates Content Credential signatures. If a certificate authority is compromised, all certificates it issued become suspect, which is why provenance systems need multiple layers of verification.

A documented trail showing every person, system, or process that handled a piece of digital evidence from its creation to its current state. In media forensics, chain of custody records help establish that evidence has not been altered and can be trusted. Digital chain of custody often relies on hash chains or Merkle trees to detect any unauthorized modifications.

An open-source audio fingerprinting library that generates compact representations of audio content based on chroma features (the distribution of energy across musical pitch classes). Used in content identification and duplicate detection. Unlike perceptual hashing for images, Chromaprint operates on the frequency content of audio signals and can match recordings even when they have been recompressed or slightly modified.

A foundation model developed by OpenAI that learns relationships between images and text descriptions. In forensics, CLIP-based detectors leverage the model's broad visual understanding to identify AI-generated images in a zero-shot manner, detecting synthetic content from generators they have never seen during training. This cross-model generalization makes CLIP a valuable component in modern detection pipelines. Related guide.

Visual or auditory distortions introduced when digital media is compressed using lossy algorithms like JPEG (images), H.264/H.265 (video), or MP3/AAC (audio). In forensics, inconsistencies in compression artifacts across different regions of an image can indicate manipulation, since edited areas may show different compression levels than the original. See: ELA, JPEG ghost detection.

The consumer-facing name for provenance information attached to content using the C2PA standard. Content Credentials typically include details about how content was created (camera, software, AI model), what edits were made, and who claims ownership. Introduced by the Content Authenticity Initiative (CAI) led by Adobe, Content Credentials are displayed as a "CR" icon that viewers can click to inspect. Full guide.

The verifiable history of a piece of digital content, covering its origin, any modifications, and its distribution path. Content provenance answers the questions: who created this, how was it made, and has it been changed? Approaches include metadata-based systems like C2PA, forensic methods that analyze the content itself, and hybrid models that combine both. Full guide.

A forensic approach that checks consistency between different types of media within the same content. For video, this means verifying that lip movements match audio, that lighting conditions are consistent between the face and background, and that audio environmental cues match the visual setting. Deepfakes often fail cross-modal checks because the synthetic component does not perfectly align with unmodified elements.

A forensic technique for identifying regions within an image that have been duplicated and repositioned, often to hide or duplicate objects. Methods include block matching, keypoint-based detection (SIFT/SURF features), and deep learning approaches. Copy-move is one of the most common forms of image manipulation and is typically analyzed alongside splice detection. Related guide.

Techniques used to expand training datasets by applying transformations like rotation, flipping, cropping, and noise addition to existing examples. In forensic detector training, data augmentation helps models generalize by exposing them to varied versions of both authentic and synthetic content. Poorly augmented training data can lead to detectors that fail when encountering real-world variations.

The complete lifecycle record of a dataset, tracking its origins, every transformation applied to it, and where it ends up. In AI contexts, data lineage maps training data sources, preprocessing steps, and downstream model outputs. Closely related to data provenance but emphasizes the transformation pipeline rather than origin verification. Full guide.

Documentation of where data comes from and how it has been processed. In forensics, data provenance helps verify that datasets have not been contaminated or manipulated. For AI training data, provenance records can address questions about licensing, consent, and representation. Standards like the W3C PROV model formalize provenance as relationships between entities, activities, and agents. Full guide.

A mathematical transform that converts spatial data into frequency components. DCT is the foundation of JPEG compression, and forensic analysis of DCT coefficients can reveal manipulation. When an image is saved as JPEG multiple times, the DCT coefficient distributions change in characteristic ways. Analyzing these patterns helps detect editing history and can identify regions that have been modified at different compression levels.

A subset of machine learning based on multi-layered neural networks that learn hierarchical feature representations from data. Deep learning powers both the creation of synthetic media (through GANs, diffusion models, and transformers) and its detection (through convolutional neural networks and foundation models). The term "deep" refers to the network's multiple processing layers, not a value judgment.

Synthetic media created using deep learning models, typically involving the replacement of one person's face, voice, or body with a generated likeness. The term originated from a Reddit username in 2017 and now broadly covers any AI-generated media designed to impersonate or mislead. Generation methods include autoencoders, GANs, and diffusion models. Detection is an active research area involving biological signal analysis, temporal coherence checks, and spectral analysis. Full guide.

A class of generative AI model that creates content by progressively removing noise from a random starting point, guided by text prompts or other conditioning inputs. Models like Stable Diffusion, DALL-E 3, Midjourney, and Imagen use this approach. Diffusion models produce different forensic artifacts than GANs, including characteristic noise patterns and texture inconsistencies at edges and fine details. Related guide.

The process of removing noise from data. In diffusion models, denoising is the core generation mechanism: the model learns to progressively remove Gaussian noise from a random input, guided by conditioning signals like text prompts. The specific denoising process used by each model leaves characteristic artifacts in the output that forensic methods can detect.

The broad discipline of recovering and investigating material found in digital devices and media. Within the context of content authenticity, digital forensics focuses on determining whether media has been created by AI, manipulated after capture, or misrepresented. Subdisciplines include image forensics, video forensics, audio forensics, and text analysis.

The recorded origin and history of a digital asset. Digital provenance provides a verifiable trail from creation through every modification and distribution step. Technologies supporting digital provenance include cryptographic hashing, blockchain-based registries, C2PA manifests, and forensic analysis. AFIP treats digital provenance as a combination of self-reported metadata and independent forensic verification. Full guide.

A forensic technique that identifies images or videos that have been compressed, edited, and recompressed. The second compression introduces detectable artifacts in the DCT coefficient distributions that differ from single-compression patterns. Localized double compression in specific image regions strongly suggests manipulation, since the edited area was saved separately before being reinserted.

Information embedded within digital content in a way that is difficult to remove without degrading the content itself. Watermarks can be invisible (imperceptible to humans) or visible (like logos overlaid on images). In content provenance, watermarks serve as a persistent signal that survives format conversion, compression, and screenshotting. See also: SynthID, robust watermark. Full guide.

A forensic technique that re-saves a JPEG image at a known quality level and compares it to the original. Regions that were recently edited or pasted in will show different error levels than the rest of the image, because they have a different compression history. ELA is one of the simplest and most widely used image forensic methods, though it has limitations with heavily compressed or multi-generation images. Related guide.

A detection system that combines multiple individual classifiers and aggregates their outputs to produce a final verdict. Ensemble approaches are common in media forensics because no single detection method works well across all types of synthetic content. By combining spectral analysis, noise pattern detection, and neural network classifiers, ensemble systems achieve higher accuracy and robustness than any individual component.

A forensic technique that matches the subtle electrical hum captured in audio or video recordings (typically 50 or 60 Hz, depending on the power grid) against a reference database of power grid frequency fluctuations. Because these fluctuations are unique to specific times and locations, ENF analysis can verify when and where a recording was made. Effective for detecting splicing or falsified timestamps. Related guide.

A measure of uncertainty or randomness in data. In text analysis, entropy quantifies how unpredictable word choices are. AI-generated text typically shows lower entropy than human writing because language models favor high-probability word sequences. Entropy analysis at the token level is one of several statistical methods used to distinguish human from machine authorship. See: perplexity.

Metadata automatically embedded in photos by cameras and smartphones, including capture date, device model, GPS coordinates, lens settings, and software used. EXIF data is valuable for provenance but easily modified or stripped during sharing and social media upload. Forensic examiners use EXIF as one data point among many, cross-referencing it with pixel-level analysis for a complete picture.

A type of deepfake in which one person's face in a video or image is replaced with another person's face. Early face swaps used autoencoders with shared encoders; modern approaches use diffusion models or GAN-based architectures. Detection methods focus on inconsistencies at face-background boundaries, unnatural temporal coherence, and mismatched biological signals.

The process of identifying and computing measurable properties from raw data that can be used by machine learning classifiers. In media forensics, features might include spectral coefficients, noise statistics, texture descriptors, or compression characteristics. The choice of features significantly affects detection accuracy. Modern deep learning detectors learn features automatically rather than relying on hand-engineered ones.

AFIP's framework for evidence-based content verification. Unlike self-reported provenance systems, FIP generates integrity assessments by analyzing the content itself using multiple forensic methods, then producing a structured report with confidence scores and supporting evidence. The protocol evaluates what the evidence shows rather than relying on claims from the content creator.

The systematic examination of digital content to determine its authenticity, origin, and modification history using scientific methods. Forensic analysis looks at the content itself rather than relying on attached metadata or labels. This includes pixel-level statistics, frequency domain inspection, compression artifact patterns, noise distribution analysis, and AI model fingerprint detection.

A large-scale AI model trained on broad data that can be adapted for many downstream tasks. In forensics, foundation models like CLIP and DINOv2 show promise as general-purpose detectors because their broad training gives them the ability to identify synthetic content from generators they have never explicitly been trained to detect. This zero-shot capability is increasingly valuable as new generators emerge.

The practice of examining media in the frequency domain rather than the spatial (pixel) domain, typically using transforms like DCT or FFT (Fast Fourier Transform). Many manipulation artifacts that are invisible in normal viewing become obvious when content is transformed into frequency space. For example, GAN fingerprints appear as distinctive spectral peaks in the high-frequency components.

A type of neural network architecture consisting of two models trained in competition: a generator that creates synthetic content and a discriminator that tries to distinguish real from generated. GANs were the dominant method for creating deepfakes before diffusion models gained popularity. Architectures like StyleGAN, ProGAN, and BigGAN produce highly realistic outputs but leave detectable GAN fingerprints.

A visible or statistically detectable flaw in content generated by a GAN. Common GAN artifacts include checkerboard patterns from upsampling operations, inconsistent eye reflections, asymmetric earring or accessory rendering, and unnatural hair-skin boundaries. While newer models reduce visible artifacts, statistical artifacts remain detectable through spectral analysis.

A subtle, model-specific pattern embedded in every image generated by a particular GAN architecture. These fingerprints arise from the upsampling operations in the generator network and appear as periodic spectral peaks when analyzed with a frequency domain transform. Each GAN architecture produces a unique fingerprint, making it possible to identify not just that an image is synthetic, but which model generated it. Related guide.

Any AI model designed to create new content such as images, video, audio, or text. Major architectures include GANs, diffusion models, autoencoders (including VAEs), and transformer-based language models. Each architecture produces different forensic signatures, which is why detection systems must account for the generator type when assessing content authenticity.

A sequence of cryptographic hashes where each entry includes the hash of the previous entry, forming a tamper-evident chain. If any link in the chain is modified, all subsequent hashes become invalid. Hash chains are used in content provenance to verify that the documented history of a file has not been altered. Related to Merkle trees and used in blockchain-based provenance systems.

A one-way mathematical function that produces a fixed-length output (hash) from any input data. Unlike perceptual hashing, cryptographic hashes (SHA-256, SHA-3) change completely if even a single bit of the input changes. This property makes them useful for verifying file integrity in chain of custody systems, but they cannot match visually similar content the way perceptual hashes can.

Software designed to modify AI-generated text so that it avoids detection by AI writing classifiers. Humanizers use techniques like synonym substitution, sentence restructuring, and style transfer to alter the statistical patterns that detectors look for. Forensic analysis of humanized text reveals persistent fingerprints, including abnormal word frequency distributions, regularized sentence lengths, and inconsistent stylometric profiles. Full guide.

The scientific analysis of digital images to detect manipulation, determine authenticity, and identify the source device or software. Methods include ELA, copy-move detection, splice detection, noise analysis, JPEG compression analysis, and AI generation detection using GAN fingerprints and diffusion artifact analysis. Full guide.

In C2PA, a reference within a content manifest that points to a source asset used during creation. If a photo is edited in Photoshop, the ingredient manifest links to the original camera capture and records what transformations were applied. This creates a traceable history of how composite works were assembled from their component parts.

A watermark embedded in digital content that cannot be perceived by humans under normal viewing conditions. Invisible watermarks encode information in the least significant bits, frequency domain coefficients, or learned latent representations of the content. Ideally, they survive common transformations like compression, cropping, and screenshotting. Examples include SynthID and various academic approaches. Related guide.

A forensic method for finding regions in a JPEG image that were saved at a different quality level than the rest of the image. When a manipulated region is pasted from another JPEG source, it carries its original compression characteristics. By re-compressing the entire image at various quality levels and computing the difference, analysts can identify "ghosts" that reveal where editing occurred. Related guide.

An ISO standard (ISO 19566-5) that defines how to embed structured metadata boxes within JPEG and other image files. C2PA uses JUMBF as its primary container format for storing manifest stores, cryptographic signatures, and assertion data. JUMBF provides a standardized way to package provenance information directly within image files. Related guide.

A forensic technique that measures the alignment between spoken audio and visible lip movements in video. Authentic speech produces consistent audio-visual synchronization with characteristic mouth shapes (visemes) matching specific phonemes. Deepfake face swaps and AI-driven lip sync tools often introduce subtle misalignment that becomes detectable through frame-by-frame analysis of mouth region movements. Related guide.

A data compression method that permanently discards some information to achieve smaller file sizes. JPEG (images), H.264/H.265 (video), and MP3/AAC (audio) are lossy formats. Forensically significant because each round of lossy compression introduces artifacts, and analysis of these artifacts can reveal editing history, recompression patterns, and splicing.

An AI system trained to predict and generate text based on patterns learned from large text corpora. Modern large language models (LLMs) like GPT-4, Claude, Gemini, and Llama generate text by predicting the next most probable token given the preceding context. The statistical predictability of this process is what makes detection possible through perplexity analysis and related methods. Related guide.

The compressed mathematical representation that generative AI models use internally to encode and manipulate content. In a diffusion model, the latent space is where the denoising process operates before the result is decoded into pixels. Understanding latent space is relevant to forensics because artifacts from the encoding and decoding process leave detectable traces in generated content, particularly at fine details and edges.

The data structure within a C2PA-enabled file that contains all provenance information. A manifest store holds one or more manifests, each describing a stage in the content's history. Manifests include assertions (claims about the content), a cryptographic signature binding them to the file, and optionally an ingredient manifest linking to source materials. Related guide.

A branch of artificial intelligence where systems learn patterns from data rather than following explicit programming rules. In forensics, machine learning powers both supervised detectors (trained on labeled real and fake examples) and unsupervised approaches that identify anomalies without labeled data. The effectiveness of ML-based detection depends heavily on training data quality, diversity, and how well it represents real-world conditions.

The overarching discipline encompassing the scientific analysis of digital media (images, video, audio, and text) to assess authenticity and detect manipulation. Media forensics combines signal processing, statistical analysis, machine learning, and domain expertise. It is distinct from metadata-based provenance because it analyzes the content directly rather than relying on attached information.

A hierarchical data structure of cryptographic hashes where each parent node is the hash of its children. Merkle trees allow efficient verification of large datasets because you can prove the integrity of any single item by checking a short chain of hashes rather than recomputing the entire dataset. Used in blockchain systems, content provenance registries, and large-scale media verification platforms.

The removal of embedded metadata from digital files, either intentionally (for privacy) or as a side effect of processing. Social media platforms routinely strip EXIF data and Content Credentials when users upload content, which is a fundamental challenge for metadata-based provenance systems like C2PA. This limitation is one reason forensic analysis, which examines the content itself, remains essential. Related guide.

A forensic approach that analyzes multiple types of media simultaneously, such as examining both the visual and audio tracks of a video. Multi-modal systems can detect inconsistencies across modalities that single-modality detectors would miss, like a deepfake video where the face looks convincing but the voice shows synthesis artifacts. This is considered the future direction for robust detection systems.

A representation of audio that captures the spectral shape of sound in a way that approximates human hearing. MFCCs are widely used as features in audio forensics, speaker verification, and voice clone detection. By analyzing MFCC patterns, forensic systems can identify inconsistencies between natural and synthetic speech, particularly in formant transitions and micro-timing variations. Related guide.

A forensic method that examines the noise patterns within digital content. Every camera sensor produces a characteristic noise pattern (known as PRNU, or Photo Response Non-Uniformity) that acts like a fingerprint. Manipulated regions disrupt this pattern, and AI-generated images produce entirely different noise distributions. Noise analysis is one of the most reliable indicators of both traditional image editing and AI generation.

A statistical technique that examines sequences of N consecutive words or characters in text. In AI text detection, n-gram distributions help distinguish human from machine writing because language models produce characteristic n-gram frequency patterns. Unusual repetition rates, missing rare n-grams, or overly smooth distributions can signal AI authorship. Often combined with perplexity and burstiness measures.

The pattern of apparent motion between consecutive video frames. In video forensics, optical flow analysis can reveal deepfake manipulation because synthetic face replacements often produce inconsistent motion patterns, particularly around face boundaries and during rapid head movements. Optical flow irregularities are closely related to temporal coherence failures.

Forensic examination of individual pixel values and their statistical properties, including color distribution, noise characteristics, and edge transitions. Pixel-level methods can detect cloning, splicing, and AI generation by identifying inconsistencies in pixel statistics that should be uniform across an authentic image. Includes techniques like ELA, noise analysis, and color filter array interpolation checks.

A technique for generating compact fingerprints of media content based on how it looks or sounds rather than its exact binary data. Unlike cryptographic hashes that change completely with any modification, perceptual hashes remain similar for visually similar content. This makes them useful for identifying copies, near-duplicates, and derivative works even after resizing, compression, or minor editing. See also: pHash, Chromaprint. Full guide.

A measure of how surprised a language model is by a given text. Low perplexity means the model could easily predict the text, which often indicates AI generation since models produce text that is, by definition, highly predictable to themselves. Human writing tends to have higher and more variable perplexity because people make creative word choices that models would not predict. Perplexity is a core feature in AI text detectors alongside burstiness. Related guide.

A specific perceptual hashing algorithm for images that generates a 64-bit hash based on the image's DCT frequency content. Two images that look similar to humans will produce similar pHash values, with the Hamming distance between hashes indicating visual similarity. Widely used in duplicate detection, reverse image search, and content identification systems. Related guide.

Any modification applied to content after initial generation or capture. In AI content, post-processing includes upscaling, color correction, inpainting, and manual touch-ups. Forensically significant because post-processing can either introduce new artifacts (creating additional detection signals) or obscure existing ones (making detection harder). Heavy post-processing is a common strategy for evading automated detection.

The documented origin and history of something. In digital contexts, provenance refers to records showing where content came from, how it was created, what modifications were made, and by whom. AFIP uses the term broadly to encompass both self-reported provenance (like C2PA metadata) and independently verified provenance (through forensic analysis). The distinction between these two approaches is central to AFIP's mission.

A unique noise pattern produced by every camera sensor due to manufacturing imperfections in individual pixels. PRNU acts as a camera fingerprint, enabling forensic analysts to link images to specific devices and detect regions where the sensor pattern has been disrupted by editing. AI-generated images lack genuine PRNU patterns entirely, making sensor noise analysis a valuable authentication tool. See: noise analysis.

Forensic analysis performed fast enough to assess content as it is being streamed, uploaded, or shared. Real-time detection is critical for video calls, live broadcasts, and social media upload pipelines where content reaches audiences within seconds. The computational demands of thorough forensic analysis make real-time performance a significant engineering challenge, often requiring lightweight models or hardware acceleration.

A technique that detects blood flow by measuring subtle color changes in facial skin visible in video. Authentic video of a person shows rhythmic skin color variations synchronized with their heartbeat. Most deepfakes fail to reproduce this signal, making rPPG analysis a powerful biological indicator of video authenticity. See: biological signal analysis.

A search technique that uses an image as the query input rather than text. Reverse image search engines use perceptual hashing and visual feature matching to find identical or similar images across the web. Useful for verifying whether an image has been previously published, identifying its original source, and detecting cases where images are presented out of their original context.

A watermark designed to survive common processing operations like compression, cropping, scaling, format conversion, and even screenshotting. Robustness is the central challenge of watermarking for content provenance, since watermarks that are easily removed provide little practical value. Modern approaches embed information in learned latent representations or spread it across frequency components to maximize survivability. Related guide.

Examination of meaning and logical coherence within content. In AI text detection, semantic analysis checks whether claims, examples, and arguments are internally consistent and factually grounded. AI-generated text sometimes produces plausible-sounding but semantically inconsistent content, particularly when making specific factual claims or providing detailed examples that do not hold up under scrutiny.

A collaborative vocabulary maintained by Google, Microsoft, Yahoo, and Yandex that defines structured data types for web content. In content provenance, Schema.org provides types like CreativeWork, MediaObject, and ImageObject with properties for author, dateCreated, and contentLocation that help establish verifiable content metadata. AFIP uses Schema.org markup to make provenance data machine-readable for search engines and verification tools.

A visual representation of the frequency content of a signal over time. In audio forensics, spectrograms reveal the energy distribution across different frequencies at each moment, making it possible to spot artifacts from AI voice synthesis, detect splicing between recordings, and identify unnatural patterns in speech that are not audible to the human ear. Related guide.

The examination of content in the frequency domain to identify patterns not visible in normal viewing or listening. For images, spectral analysis using FFT reveals GAN fingerprints and manipulation artifacts. For audio, it uncovers synthesis patterns, compression boundaries, and formant inconsistencies. Spectral methods are among the most reliable forensic techniques because generators struggle to produce natural frequency distributions.

The process of determining which device, software, or AI model produced a specific piece of content. Source attribution can use camera sensor fingerprints (PRNU), software metadata, GAN fingerprints, or model-specific noise patterns to link content back to its origin. Particularly valuable in legal and journalistic contexts where establishing the source of evidence matters.

A forensic technique for identifying regions of an image that have been inserted from a different source. Unlike copy-move detection (which finds duplicated regions within the same image), splice detection looks for inconsistencies in noise patterns, lighting direction, compression artifacts, and color characteristics that indicate content from multiple sources has been combined. Related guide.

Google DeepMind's watermarking technology for AI-generated content, initially deployed for images created by Imagen and later extended to text (via Gemini) and audio. SynthID embeds imperceptible signals during the generation process rather than adding them after the fact, making the watermark tightly integrated with the model's output. Currently used across Google's AI products. Related guide.

The practice of hiding information within other content so that the existence of the hidden message is not apparent. In media forensics, steganographic analysis detects hidden data embedded in images, audio, or video. Steganography differs from watermarking in purpose: watermarks identify ownership, while steganography conceals the very fact that a message exists. Both can serve provenance functions.

A machine learning approach where models are trained on labeled examples (such as images labeled "real" or "fake"). Most current forensic detectors use supervised learning, which means their accuracy depends on the quality and diversity of their training datasets. A supervised detector trained only on GAN outputs may struggle with diffusion model content, highlighting the need for diverse training data or zero-shot approaches.

The statistical analysis of writing style to determine authorship or detect AI generation. Stylometric features include vocabulary richness, sentence structure distribution, punctuation habits, function word frequencies, and syntactic complexity patterns. In AI text detection, stylometry identifies inconsistencies between a claimed author's historical writing patterns and AI-generated text that mimics them. Related guide.

AI technology that converts written text into spoken audio. Modern neural TTS systems (like Bark, XTTS, and ElevenLabs) produce highly natural-sounding speech that is increasingly difficult to distinguish from human recordings. Forensic detection of TTS output focuses on MFCC anomalies, formant transitions, breathing pattern absence, and prosodic irregularities. Related guide.

The consistency of visual elements across sequential video frames. Authentic video maintains natural continuity in lighting, shadows, reflections, and object positions from one frame to the next. Deepfake generation often introduces temporal artifacts like flickering at face boundaries, inconsistent lighting across frames, and unnatural motion in hair and accessories. Analyzing temporal coherence is a primary method for detecting video-based deepfakes. Related guide.

A hierarchical structure of trusted entities, where each entity vouches for the next, anchored by a root of trust (typically a certificate authority). In C2PA, the trust chain links a content signature back through intermediate certificates to a trusted root, allowing verifiers to confirm that the signing entity is legitimate. The integrity of provenance systems depends on the strength and accessibility of their trust chains.

A technique where a model trained on one task is adapted for a different but related task. In forensics, transfer learning enables detectors trained on one generator's output to partially detect content from other generators. Foundation models leverage transfer learning extensively, using their broad pretraining knowledge to generalize across detection scenarios without needing generator-specific training data.

The ratio of unique words (types) to total words (tokens) in a text, used as a measure of vocabulary diversity. Human writing typically shows higher variation in type-token ratio across different passages, while AI text tends toward a narrower band. When combined with perplexity and burstiness, type-token ratio helps distinguish human from machine-generated writing. Related guide.

A generative model that learns a probabilistic mapping between input data and a structured latent space. VAEs are used as components in diffusion model pipelines (encoding images to latent space for processing) and produce characteristic artifacts at fine details and sharp edges. These VAE artifacts are forensically distinct from those produced by GANs, aiding in generator identification.

The scientific analysis of video content to detect deepfakes, manipulation, and AI generation. Video forensics combines frame-by-frame image analysis with temporal methods that examine consistency across frames. Techniques include temporal coherence analysis, audio-visual synchronization checking, biological signal analysis, and motion trajectory verification. Full guide.

The use of AI to generate synthetic speech that mimics a specific person's voice. Modern voice cloning systems can produce convincing results from as little as three seconds of reference audio. Detection methods analyze MFCC patterns, formant consistency, breathing markers, and micro-timing variations that synthetic voices struggle to reproduce naturally. Voice cloning fraud has driven rapid growth in audio forensic verification. Related guide.

A mathematical representation of the unique acoustic characteristics of an individual's voice, including fundamental frequency, formant positions, vocal tract resonance, and speaking rhythm. Voiceprints are used in speaker verification and can help detect voice clones by comparing the synthetic output against a known reference. However, advanced cloning systems are increasingly able to approximate voiceprint characteristics.

The process of extracting or verifying a digital watermark embedded in content. Detection can be informed (using a secret key to extract the watermark) or blind (detecting the presence of a watermark without knowing the original). In content provenance, watermark detection allows platforms and users to verify whether AI-generated content carries its origin signal, even after the content has been shared, compressed, or modified. Related guide.

The ability of a forensic classifier to identify AI-generated content from a model it has never been trained on. Traditional detectors are trained on outputs from known generators and may fail when encountering new ones. Zero-shot approaches, often built on foundation models like CLIP, learn general features of synthetic content that transfer across generators. This capability is increasingly important as new AI models emerge faster than detectors can be retrained. Related guide.