Technology
Technology OverviewImage Trust LayerDocument Trust LayerAudio Trust LayerInformation Trust LayerSource VerificationProducts
Fraud ScannerSafeCamSafeMicGardioIndustries
InsuranceBanking & KYCAuthoritiesMedia & PressHR & ExpensesDefenseResources
BlogCase StudiesAPI DocsPricingPress AreaCompany
About UsJobsContactThe Image Trust Layer is VAARHAFT's proprietary deep learning architecture for detecting deepfakes, AI-generated images, and image manipulation. Trained on real-world data, not academic lab benchmarks.
Each module handles a specific image analysis task. Proprietary neural networks for AI detection, deterministic algorithms for matching, and deep metadata extraction. All accessible through a single endpoint.
Detects fully AI-generated images and deepfakes: GANs, diffusion models, neural renderers. Standalone, proprietary neural network.
Detects post-capture manipulations on real photos: deepfake face swaps, inpainting, splicing. Standalone, proprietary neural network.
Perspective-invariant matching beyond SHA256. 100% deterministic, no neural network embeddings.
Pixel-level attribution maps that visualize exactly which regions triggered detection. Full transparency for human review.
Extracts and cross-references EXIF, XMP, and C2PA metadata to verify provenance, device origin, and editing history.
No off-the-shelf foundation models. Our entire detection pipeline is built from scratch with hierarchical spatial architectures, trained exclusively on proprietary forensic datasets.
Context-Aware Anomaly Detection
Detects anomalies in spatial context not just isolated pixel artifacts.
Intrinsic Explainability
Explainability is an inherent mathematical part of backpropagation not generated post-hoc like GradCAM or ELA. The proof is part of the classification.
Deterministic, No LLM
100% reproducible results. No stochastic LLM, no prompt engineering, no hallucinations.
Adversarial Hardening
Hardened against FGSM and PGD adversarial perturbations as well as real-life data laundering (WhatsApp, Telegram, screenshots).
A highly optimized, multi-stage feature extraction pipeline trained to detect manipulation-specific artifacts, even after severe messenger compression. Intrinsic explainability as a mathematical byproduct no post-hoc XAI.
Academic benchmarks test on clean, high-resolution original images. Fraudsters almost never use these. The benchmarks from the majority of image forensics providers are worthless.
Our model was not trained in a lab but on real-world data. We evaluate our metrics explicitly on "washed" data (data laundering).
82% Precision @ 1% Prevalence: Even when 99 out of 100 images are real, our model is correct 82% of the time when flagging an image as "fake".
Our performance holds even when an AI image was:
FP-Catcher Module
Pure image analysis is not enough for production. Specialized false-positive catcher modules are adapted to real production conditions. Example: Logos are automatically detected and filtered out in the first step, as they would otherwise lead to duplicates or false hits in reverse image searches.
Different use cases require different trade-offs. Automated rejection needs maximum precision. Screening with human-in-the-loop maximizes recall.
Choose the mode that matches your risk appetite or let us calibrate your model individually.
Balanced: Balanced trade-off between precision and recall. Default configuration for most enterprise customers.
Our out-of-the-box metrics are industry-leading, but every company has unique data distributions. We offer deep system calibration to tune the model precisely to our enterprise customers' requirements.
Through advanced probability calibration and model finetuning, we ensure that a confidence score of 90% truly corresponds to a 90% probability of a fake optimal for seamless integration into existing risk engines.
Our model was not trained in an academic lab. Over 10 million in-house curated data points hand-edited by image editing experts simulate extreme real-world conditions including massive messenger compression and targeted image degradation.
Most open-source detectors and academic models are trained on clean, high-resolution originals from public datasets like DFFD or FaceForensics++. These benchmarks do not reflect production reality: real fraud images arrive compressed, screenshotted, cropped, and shared through multiple messengers. Our proprietary training data mirrors exactly these conditions, which is why our detection accuracy holds where academic models collapse.
Zero-Day GenAI Defense
When new AI models or versions of them are released, we integrate their signatures within days not weeks.
GPU clusters (NVIDIA Tensor Core) scale dynamically. 10 requests at night or 5,000 RPM during a campaign latency remains stable.
Specialized sub-models vote independently on generation and manipulation probability, combined via calibrated ensemble.
Dedicated endpoints no serverless cold-start penalty. Inference optimized via TensorRT (INT8 Quantization).
Average end-to-end latency including ensemble voting, heatmap generation, and JSON response.
Zero-retention API: We don't store any images. RAM is cleared immediately after inference. Privacy by design.
Supports all common image formats natively: JPEG, PNG, TIFF, HEIC, WebP. No conversion, no pre-processing required.
Direct upload up to 10 MB for single images and compact PDFs. For large payloads and batch pipelines, asynchronous S3 offloading up to 80 MB with webhook notification on completion.
Over 10 million in-house curated data points simulating extreme real-world conditions including massive messenger compression and targeted image degradation. Robustness where conventional SOTA models break down.
When new architectures like Midjourney v7 or Sora are released, we integrate their signatures within days. Radical agility in the cat-and-mouse game of AI generators.
Developed and hosted in Germany. Zero-retention API: RAM is cleared immediately after inference, no customer data stored. Fully auditable, EU AI Act compliant, no third-party data leakage.
Our architecture generates explainability not after the fact, but as a mandatory mathematical byproduct. Granular JSON data and spatial heatmaps prove pixel-precisely why an image was flagged.
Built from the ground up for scale. Asynchronous webhooks, deterministic processing without LLM overhead, and seamless S3 payload offloading for gigabyte-heavy document pipelines.
Specialized modules for real production conditions. Logo detection, duplicate filtering, and adaptive threshold adjustment minimize false positives.
The Image Trust Layer is the core detection engine of the VAARHAFT Fraud Scanner available as a fully managed API. No infrastructure, no model maintenance. Integrate once, benefit permanently.
{
"suspicion_level": "Green",
"Files": {
"photo_01.jpg": { … },
"photo_02.jpg": { … }
},
"caseNumber": "Case 123A",
"sessionId": "5a8d8fd2-3317-4964-8298-caa96cd5cfa3",
"modelVersions": {
"imagesModelVersion": "vh-mo-images-alpha",
"documentsModelVersion": "vh-mo-docs-alpha"
},
"tokensConsumed": 12,
"attachments": {
"heatmaps": [ … ],
"analysis_report": [ … ]
}
}
Open-source tools like FaceForensics++ detectors or DIRE are trained on academic datasets with clean, high-resolution originals. They perform well in lab conditions but collapse when images are compressed, screenshotted, or shared via messengers. Our models are trained exclusively on proprietary real-world data, including washed, degraded, and adversarially manipulated images, and deliver production-grade accuracy where academic models fail.
We detect images from all major GenAI models, including (but not limited to) Stable Diffusion, Dall-E / ChatGPT, Flux, Nano Banana.
We deploy specialized FP-Catcher modules adapted to real production conditions. For example, logos and watermarks are automatically detected and filtered in the first step, as they would otherwise trigger false hits. Additionally, enterprise customers can configure custom threshold tuning and probability calibration for their specific data distributions.
Our models are hardened against known adversarial perturbation techniques including FGSM and PGD attacks, as well as real-world data laundering strategies like multiple rounds of messenger compression, screenshots, and print-scan cycles. We continuously test against new evasion techniques and update our defenses accordingly.
None. The Image Trust Layer uses a zero-retention architecture: images are processed in GPU RAM and cleared immediately after inference. No customer images are stored, logged, or retained. The API is fully stateless, EU AI Act compliant, and hosted in Germany.
Average end-to-end latency is under 5 seconds including ensemble voting, heatmap generation, and JSON response. Our GPU clusters scale dynamically, from 10 requests at night to 5,000 RPM during campaigns, with zero cold-start penalty thanks to dedicated TensorRT-optimized endpoints.
Every VAARHAFT product is built on the same proprietary technology base: five specialized trust layers covering images, documents, audio, contextual information, and capture-time source verification. Combined, they form the technological foundation of the VAARHAFT Trust Suite.
Detects deepfakes, AI-generated images, and image manipulation.
Book a demo and see how our detection performs on your real-world pipeline. We'll walk you through integration, threshold configuration, and deployment options.
