Temporal Image Forensics

Updated 10 July 2026

Temporal image forensics is the field of verifying when a visual asset was captured by analyzing time-dependent physical and digital traces.
It employs methods like sun-position validation, sensor defect tracking, and frequency-domain analysis to detect inconsistencies and forged temporal evolution.
The research also focuses on localizing manipulated segments and proactively tracing motion in image-to-video generation, addressing content bias and realistic deployment challenges.

Temporal image forensics encompasses methods that verify or infer the temporal properties of visual evidence: when an image was acquired, whether a claimed timestamp is consistent with scene content and location, whether a video exhibits forged temporal evolution, and, in proactive settings, whether the original source image can be recovered from generated motion. In the narrower acquisition-pipeline sense, it is the science of estimating the age of a digital image from time-dependent traces; in broader multimedia forensics, it includes timestamp verification, temporal deepfake detection, temporal forgery localization, and motion-aware tracing in image-to-video generation (Jöchl et al., 9 Sep 2025, Padilha et al., 2021, Kim et al., 3 Jul 2025, Chen et al., 16 Apr 2026).

1. Scope, tasks, and forensic settings

The field contains several distinct but related problem formulations. In relative image-age estimation, the traditional setting assumes a trusted set of chronologically ordered images from the same camera and an untrusted set of images from the same camera with unknown capture times; the task is to place each untrusted image into the temporal axis defined by the trusted images (Jöchl et al., 9 Sep 2025). The same review also proposes a new setting in which the camera itself is available, calibration images can be captured at seizure time, and the goal can extend from relative ordering toward absolute age estimation (Jöchl et al., 9 Sep 2025).

A broader operational view arises in misinformation and multimedia forgery. One branch verifies contextual claims such as time and location for outdoor imagery, either by comparing shadow-inferred sun position with astronomically computed sun position from claimed metadata (Li et al., 2018) or by learning the consistency of image content, timestamp, geographic location, and overhead imagery through a supervised model $P(y \mid G,t,l,S)$ (Padilha et al., 2021). A second branch studies video forgery by modeling temporal inconsistencies in frequency space, either through spatial–temporal DCT features and temporal attention (Wang et al., 2022) or through pixel-wise temporal FFT and localized attention over artifact-prone regions (Kim et al., 3 Jul 2025). A third branch treats partial manipulation as temporal forgery localization in untrimmed sequences, where the goal is to locate all forged segments rather than assign a single real/fake label (Yin et al., 10 Jun 2025). A fourth, proactive branch embeds a learnable forensic template into a source image so that later image-to-video generations can be traced backward through motion fields (Chen et al., 16 Apr 2026).

Problem family	Core cue	Representative paper
Image age approximation	Time-dependent traces from the acquisition pipeline	(Jöchl et al., 9 Sep 2025)
Context verification	Sun position or learned geo-temporal plausibility	(Li et al., 2018, Padilha et al., 2021)
Video deepfake detection	Spatial–temporal or pixel-wise temporal frequency anomalies	(Wang et al., 2022, Kim et al., 3 Jul 2025)
Temporal forgery localization	Instant-level anomaly relative to sequence context	(Yin et al., 10 Jun 2025)
Proactive I2V forensics	Motion-synchronized embedded template and flow reversal	(Chen et al., 16 Apr 2026)

This diversity implies that “temporal” is not limited to dating. It can denote capture-time verification, temporal consistency, temporal localization of manipulated segments, or the tracing of evidence through generated motion.

2. Contextual verification of claimed time and place

A physically grounded line of work validates contextual information in outdoor images by exploiting the law of nature that sun position varies with the time and location. “IMAGEGUARD” estimates sun altitude and azimuth from a single vertical object and its shadow, while allowing the camera to be tilted up or down; it then compares that shadow-inferred sun position with the sun position computed from claimed capture time and location using astronomical algorithms (Li et al., 2018). The paper defines altitude difference $d_h$ , azimuth difference $d_A$ , and sun position distance $d_p$ , and reports that, by setting the thresholds to be $9.4$ degrees and $5$ degrees for the sun position distance and the altitude angle distance, respectively, the system can correctly identify $91.5\%$ of falsified photos with fake contextual information (Li et al., 2018).

This physics-based approach is explicitly tailored to outdoor, daylight images with visible shadows, approximately level ground, and approximately vertical objects. Its temporal resolution depends on solar geometry: the paper reports detectable time-of-day falsifications of at least $16$–$40$ minutes depending on season and base time, detectable date shifts of at least $13$– $d_h$ 0 days, and the ability to detect location shifts greater than approximately $d_h$ 1 miles from New York at fixed time and date (Li et al., 2018). The method is not applicable when the sun is below the horizon, under heavily diffused lighting, or when the shadow geometry assumptions break.

A complementary, data-driven paradigm treats timestamp verification as content-aware supervised consistency. The formulation takes a ground-level outdoor image $d_h$ 2, a claimed timestamp $d_h$ 3, a location $d_h$ 4, and an overhead/basemap image $d_h$ 5, and learns $d_h$ 6, where $d_h$ 7 denotes consistency and $d_h$ 8 inconsistency (Padilha et al., 2021). The architecture uses four encoders for $d_h$ 9, $d_A$ 0, $d_A$ 1, and $d_A$ 2; predicts a two-class consistency output; and adds two auxiliary transient-attribute heads that estimate 40 attributes such as snow, rainy, foggy, sunrise, sunset, night, summer, autumn, lush, beautiful, and stressful (Padilha et al., 2021). These auxiliary tasks are both regularizers and explanation channels, because disagreements between $d_A$ 3 and $d_A$ 4 reveal which time-sensitive attributes conflict.

On the Cross-View Time dataset, this content-aware model improves timestamp tampering detection accuracy from $d_A$ 5 to $d_A$ 6, with AUC $d_A$ 7 for the best DenseNet model with transient attributes on the shared-camera protocol (Padilha et al., 2021). The ablations show that time alone with the image is insufficient; adding location yields the largest single boost, satellite imagery adds complementary structural context, and auxiliary transient tasks modestly but consistently improve performance (Padilha et al., 2021). Under the more realistic camera-disjoint split, the same model reaches $d_A$ 8 accuracy and AUC $d_A$ 9, still above the Salem et al. baseline at $d_p$ 0 accuracy and AUC $d_p$ 1 (Padilha et al., 2021). The same verification model can also estimate missing time-of-capture by evaluating all $d_p$ 2 month-hour candidates and visualizing a consistency heatmap over the month $d_p$ 3 hour grid (Padilha et al., 2021).

Taken together, these two directions illustrate a central split in temporal image forensics: explicit physical validation versus learned geo-temporal plausibility. The former provides a narrow but interpretable consistency check; the latter broadens applicability by learning global appearance cues without relying on explicit physics models or external weather data.

3. Acquisition-pipeline age traces and image age approximation

In the acquisition-pipeline formulation, the principal temporal evidence consists of age traces: subtle, time-dependent artifacts introduced by the camera hardware and processing chain. The review identifies two characterized traces, in-field sensor defects and sensor dust, and emphasizes that most work addresses relative rather than absolute age (Jöchl et al., 9 Sep 2025).

In-field sensor defects are largely hot pixels and partially-stuck hot pixels that develop after manufacturing. Their physical origin is associated primarily with terrestrial cosmic rays, especially neutrons, which cause displacement damage in the silicon bulk; empirically, defects appear instantaneously between captures, then largely persist, and the number of defects grows approximately linearly in time (Jöchl et al., 9 Sep 2025). The review summarizes reported growth rates such as approximately $d_p$ 4 defects/month/Mpix or $d_p$ 5 defects/1000 images/Mpix from Dudas et al., and notes that inter-defect times follow an exponential distribution, consistent with a Poisson process of constant rate (Jöchl et al., 9 Sep 2025). A widely used sensor-output model writes

$d_p$ 6

where $d_p$ 7 is the ideal noiseless image, $d_p$ 8 is PRNU, $d_p$ 9 is dark current, $9.4$0 is a fixed offset, $9.4$1 collects time-dependent settings, and $9.4$2 is other noise (Jöchl et al., 2023, Jöchl et al., 9 Sep 2025).

Sensor dust is the second established age trace, especially for interchangeable-lens systems. Dust particles on or near the protective cover glass cast blurred spots whose visibility depends strongly on aperture; their location on the sensor is stable until cleaning, and their apparent image position shifts with focal length (Jöchl et al., 9 Sep 2025). The review treats dust as a plausible age trace whose monotonicity is weaker than that of hot pixels because cleaning events can reset the pattern (Jöchl et al., 9 Sep 2025).

Classical temporal image forensics methods model these traces explicitly. The review describes information-theoretic temporal ordering using PRNU clusters, maximum-likelihood dating from defect onset, and machine-learning classification on residuals at known defect locations (Jöchl et al., 9 Sep 2025). Among these, methods that only use residuals at fixed defect coordinates are presented as more robust against content bias because they do not permit the model to exploit arbitrary scene content (Jöchl et al., 9 Sep 2025).

A major controversy concerns deep learning. The paper on content bias argues that in temporal image forensics it is not evident that a neural network trained on images from different time-slots exploits solely image age related features; images taken in close temporal proximity can share common content properties, and a network may exploit those instead (Jöchl et al., 2023). To test this, the paper introduces average-image probes: class-wise average images $9.4$3, average color images $9.4$4, structural residual images $9.4$5, and median-filtered average images $9.4$6 (Jöchl et al., 2023). On synthetic data with a controlled embedded age signal sampled from dark-field images, the probes behave as expected for genuine age-signal usage; on real image-age classification, the same analysis shows that a deep learning approach proposed in the context of age classification is most likely highly dependent on the image content (Jöchl et al., 2023).

The 2025 review generalizes this criticism. It argues that CNN-based dating methods on uncontrolled real-scene datasets can achieve high classification accuracy while relying primarily on scene content, colors, or incidental biases, and that eXplainable Artificial Intelligence methods such as Grad-CAM++, Score-CAM, average-image tests, and masking experiments are indispensable for verifying that a model uses genuine age traces rather than dataset artifacts (Jöchl et al., 9 Sep 2025). It also reports that a CNN dating palmprint images does appear to rely on a scanner dust or dirt pattern, which is a valid acquisition-related age trace, but under conditions far more controlled than natural-image dating (Jöchl et al., 9 Sep 2025).

4. Temporal inconsistency in video: frequency-domain approaches

For video deepfake detection, a major transition has been from frame-wise spatial analysis toward explicit temporal modeling in the frequency domain. One approach, FCAN-DCT, treats forgery clues as spatial–temporal frequency patterns that fluctuate across frames (Wang et al., 2022). It begins with uniformly sampled face crops, extracts per-frame backbone feature maps, applies a channel-wise 2D-DCT, enhances medium and high frequencies by a weighting rule, partitions the spectrum into blocks, and uses max-pooling to obtain a compact spatial frequency descriptor. A Frequency Temporal Attention module then computes per-frame attention over frequency bands, and a weighted aggregation produces a video-level forgery feature (Wang et al., 2022). The temporal signal is therefore not 3D motion in RGB space, but instability of DCT-band saliency across frames.

On WildDeepfake, FCAN-DCT with ResNet50 reaches $9.4$7 accuracy and $9.4$8 AUC, improving over Xception at $9.4$9 accuracy and $5$0 AUC, F$5$1-Net at $5$2, and PEL at $5$3 (Wang et al., 2022). On Celeb-DF (v2), FCAN-DCT with ResNet50 reaches $5$4 accuracy and $5$5 AUC, exceeding or matching strong 3D-CNN baselines (Wang et al., 2022). Cross-dataset results are also strong: when training on FF++ and testing on Celeb-DF (v2), FCAN-DCT with Xception attains $5$6 AUC versus $5$7 for the Xception baseline; when training on WildDeepfake and testing on Celeb-DF, it reaches $5$8 accuracy, surpassing a temporal RGB method denoted DAM at $5$9 (Wang et al., 2022). The same pipeline also generalizes to near-infrared data through DeepfakeNIR, the first video forgery dataset on near-infrared modality (Wang et al., 2022).

A later line of work argues that even spatial–temporal frequency methods still overlook a more direct cue: how each pixel oscillates over time. The pixel-wise temporal frequency approach performs a 1D Fourier transform along the time axis for each pixel after median-filter-based preprocessing and grayscale conversion, producing a temporal frequency vector for every spatial location (Kim et al., 3 Jul 2025). A global temporal-frequency volume $91.5\%$ 0 is processed by a 2D ResNet-50, an Attention Proposal Module learns five discriminative local parts, a feature blender injects temporal-frequency features into a frozen 3D ResNet-50, and a Joint Transformer Module combines a Spatial Transformer Encoder and a Temporal Transformer Encoder (Kim et al., 3 Jul 2025). The explicit target is local flicker, unstable motion or shape, inconsistent facial dynamics, and subtle temporal artifacts that can be diluted by global spatial spectra.

This method reports an average AUC of $91.5\%$ 1 over Celeb-DF, DFDC, FaceShifter, DeeperForensics, and DFD when trained on FF++, outperforming or matching state-of-the-art temporal baselines; on KoDF it reaches AUC $91.5\%$ 2; and in leave-one-synthesis-out experiments on FF++ it reports AUC values of $91.5\%$ 3 for DF, $91.5\%$ 4 for FS, $91.5\%$ 5 for F2F, and $91.5\%$ 6 for NT (Kim et al., 3 Jul 2025). It also provides a strong shuffling test: when frame order is randomly permuted, detection performance on FF++ drops from $91.5\%$ 7 to $91.5\%$ 8 AUC, indicating that the detector depends on coherent temporal dynamics rather than static appearance alone (Kim et al., 3 Jul 2025). The main reported failure mode is heavy compression, which attenuates high-frequency temporal content and therefore weakens the target signal (Kim et al., 3 Jul 2025).

These two approaches define complementary versions of temporal frequency forensics. FCAN-DCT treats the temporal clue as instability of frequency-band responses across frames, whereas the pixel-wise method treats each pixel’s temporal spectrum as the primary forensic object. Both are attempts to move beyond per-frame artifact detection toward explicitly modeled temporal inconsistency.

5. Segment localization and proactive temporal tracing

Temporal image forensics increasingly distinguishes between global detection and localization. UniCaCLF formulates temporal forgery localization as anomaly detection over instant-level features $91.5\%$ 9, where forged intervals are sparse relative to a mostly genuine sequence (Yin et al., 10 Jun 2025). The framework builds a global context $16$0, refines instant features through a context-aware perception layer, and trains a context-aware contrastive loss that pulls genuine instants toward the context while pushing forged instants away (Yin et al., 10 Jun 2025). The Heterogeneous Activation Operation boosts instants that are far from context according to negative cosine similarity, while the Adaptive Context Updater re-estimates the context by weighting temporally aligned, context-similar instants more heavily (Yin et al., 10 Jun 2025). The classification and regression heads then output forged segments with temporal boundaries.

The design is universal in the precise sense used by the paper: the same framework is applied to video-only, audio-only, and audio-visual features, using pre-trained backbones such as TSN, I3D, ResNet50, BYOL-A, and wav2vec (Yin et al., 10 Jun 2025). On LAV-DF, UniCaCLF reports [email protected] of $16$1 and average AP $16$2, improving markedly over ActionFormer, TriDet, BA-TFD+, AVTFD, UMMAFormer, and MFMS at this strict IoU level (Yin et al., 10 Jun 2025). On AV-Deepfake1M it reaches [email protected] $16$3 and average AP $16$4; on TVIL it reaches [email protected] $16$5 and average AP $16$6; and on HAD it reaches [email protected] $16$7 (Yin et al., 10 Jun 2025). The cross-dataset test from HAD to Psynd is especially significant: average AP rises to $16$8, above TriDet at $16$9 and ActionFormer at $40$0, which the paper attributes to sample-by-sample context-aware contrastive coding that suppresses cross-sample influence (Yin et al., 10 Jun 2025).

A separate direction redefines temporal forensics for image-to-video generation. “Flow of Truth” is presented as the first proactive framework focusing on temporal forensics in I2V generation (Chen et al., 16 Apr 2026). Instead of asking where a static frame was tampered, it asks how a source image $40$1 evolved into video frames $40$2, and formalizes the forensic goal as learning

$40$3

followed by a fusion operator

$40$4

so that the original source can be reconstructed even if early frames are dropped (Chen et al., 16 Apr 2026). The core conceptual shift is to redefine video generation as the motion of pixels through time rather than the synthesis of independent frames (Chen et al., 16 Apr 2026).

The method embeds a learnable invisible template $40$5 into the source image, decodes a motion-aware template from simulated generated frames, estimates dense flow in template space rather than RGB space, and reverses motion through warping and confidence-weighted fusion (Chen et al., 16 Apr 2026). It is proactive because it requires the source image to be protected in advance. On the I2V benchmark comprising 816 videos and 69K frames across CogVideoX, Wan2.2, Kling2.1, and Dreamina S2.0, FoT improves truth recovery over raw generated frames. For example, in the frame-level setting for the s10–40 pixels motion range, “Forged” yields PSNR $40$6, SSIM $40$7, LPIPS $40$8, CLIP-S $40$9, and DINO-S $13$0, whereas “FoT” yields PSNR $13$1, SSIM $13$2, LPIPS $13$3, CLIP-S $13$4, and DINO-S $13$5 (Chen et al., 16 Apr 2026). Under a $13$6 frame-dropping attack, “Forged” yields PSNR $13$7 and SSIM $13$8, while “FoT” yields PSNR $13$9 and SSIM $d_h$ 00 (Chen et al., 16 Apr 2026). The same motion-aware features also support tampering localization and improve the robustness of other watermarking systems under I2V transformations (Chen et al., 16 Apr 2026).

Together, UniCaCLF and FoT show that temporal forensics is no longer confined to deciding whether a sequence is fake. It increasingly seeks either precise temporal boundaries for manipulated segments or explicit recovery of the original source behind generated motion.

6. Reliability, failure modes, and research directions

A recurrent pattern across the literature is that temporal evidence is often weak, local, and confounded by nuisance factors. In timestamp verification, very small changes such as $d_h$ 01 month or $d_h$ 02 hour are difficult, with detection rate below $d_h$ 03, whereas $d_h$ 04 or $d_h$ 05 yields more than $d_h$ 06 detection; night-time scenes within $d_h$ 07– $d_h$ 08 are visually similar, and large location perturbations can turn correct timestamps into false positives unless location augmentation is used (Padilha et al., 2021). In sun-position validation, applicability is limited by visible direct-light shadows, accurate camera orientation, and approximate scene geometry (Li et al., 2018). In frequency-based video detection, severe compression attenuates the target signal, and specular reflections on glasses can produce false positives (Kim et al., 3 Jul 2025). In FCAN-DCT, robustness to frame-rate changes, heavy re-encoding, more aggressive temporal editing, and adversarial attacks remains to be explored (Wang et al., 2022). In Flow of Truth, performance drops for large and complex motion, and the method requires prior embedding, so it does not address non-cooperative content (Chen et al., 16 Apr 2026). UniCaCLF, while strong, remains fully supervised and depends on the quality of the underlying backbones (Yin et al., 10 Jun 2025).

The most consequential methodological controversy is content bias. Both the dedicated study on image age approximation and the later review argue that high accuracy on real-image temporal classification can be misleading when scene content, lighting, camera settings, or incidental acquisition changes correlate with time labels (Jöchl et al., 2023, Jöchl et al., 9 Sep 2025). The recommended response is not only better modeling but stronger diagnosis: average-image probes, CAM-based saliency, masking experiments, standardized scene acquisition, synthetic datasets with controlled age signals, and calibration images such as dark-field or bright-field captures (Jöchl et al., 2023, Jöchl et al., 9 Sep 2025). In this sense, explainability is not peripheral; it is part of the evidential validation of a temporal forensic method.

Several future directions are already explicit in the cited work. Timestamp verification can be extended beyond outdoor scenes, integrated with explicit sun-position or weather priors, and combined with textual claims, posting time, and platform-level metadata (Padilha et al., 2021). Pixel-wise temporal frequency could be fused with optical flow or physiological signals and made compression-aware (Kim et al., 3 Jul 2025). UniCaCLF suggests weakly supervised or unsupervised temporal forgery localization and broader application to frame insertion, deletion, or speed manipulation (Yin et al., 10 Jun 2025). FoT suggests better motion modeling, extension to text-to-video and video-to-video, joint training with watermarking, and model attribution (Chen et al., 16 Apr 2026). The review of age-trace methods argues for more realistic forensic settings in which the camera is available, calibration images are acquired at seizure time, and absolute age estimation is attempted with physics-informed models of defect and dust evolution (Jöchl et al., 9 Sep 2025).

A plausible implication is that temporal image forensics is becoming a family of methods for joint reasoning over content, time, place, motion, and provenance rather than a single task. Its strongest results arise when temporal cues are modeled explicitly, tied to interpretable physical or signal-level mechanisms, and stress-tested against content bias and deployment conditions.