Deepfake Detection Platform for Enterprise

What is identifAI?

IdentifAI is an enterprise-level deep fake detection platform developed by Italian company IdentifAI Labs S.P.A.; it is used to determine whether images, videos, and audio clips were created by humans or generated/modified using artificial intelligence.

The product views media forensics as a dynamic issue within cybersecurity, offering web analysis, REST APIs, a remote MCP Server, video conference agents, and options for enterprise deployment. The detection results are used for risk screening and manual verification; they do not constitute conclusive proof of the authenticity of the content, the identities of the individuals involved, or any legal facts.

Main functions

• Image detection: Upload a file or provide a link to an accessible image, and you will receive information regarding whether the image represents a human or an object, along with a confidence score, the results from various models, and a comprehensive heuristic assessment.
• Video detection: 5, 10, or 30 frames are extracted for analysis; options include key frame selection, face swapping, detection of alterations, audio track analysis, and identification of unsafe content, with the overall findings then compiled.
• Voice detection: Analyzes recordings or online audio to identify signs of synthetic voices and voice cloning; suitable for call centers, telephone authentication, and audio verification.
• Detection of face and identity-related attacks: The REVELIO FACE series is designed to deal with deepfaked selfies, synthetic avatars, camera intrusions, and identity fraud; it can be integrated into KYC and account opening processes.
• Tampering and heat maps: It is possible to conduct additional checks for signs of traditional editing or stitching, and heat maps can be used to identify areas in an image that are considered to be artificially created; these additional analyses result in a consumption of credits.
• Real-time meeting Agent: It continuously analyzes the video and audio of meetings, and enables the detection of imposters through configuration via a unified portal, as well as monitoring and report viewing.
• Manual correction and tracking: The API allows for overriding classification results based on file hashes, associating them with business reference numbers, retrieving results in batches, and checking the remaining credit.

Three types of detection models

The video is broken down into image frames and optional audio tracks; each of these components is analyzed by dedicated models, after which a heuristic engine generates an overall assessment. The older models can still be used to support earlier versions of the generator, but their capabilities should not be confused with those of the current main model.

How to understand performance metrics?

These metrics come from benchmarks that include both academic and proprietary samples, and they do not guarantee the same performance in any other context, with any new generator, or under any different compression conditions. Companies should use their own actual samples to evaluate false positives, false negatives, and threshold values before deciding on automated actions to take.

Web usage process

1. Register an account, verify your email address and log in, then check your current credit balance as well as the models that can be used with your account.
2. You can upload images, videos, or recordings, as well as provide media URLs; it is advisable to use materials that are as close as possible to the original files and for which you have the legal right to use them.
3. Select the desired analysis model as well as additional options such as face detection, detection of alterations, heatmaps, audio tracks, or unsafe content.
4. After confirming the selection, submit the task; wait for the asynchronous analysis to be completed, and then view the comprehensive assessment, model scores, video frames, and available heatmaps.
5. Combine the results with business data, identity verification, metadata, and manual review; do not reject customers, remove content, or conclude that fraud has occurred based on a single score alone.

How REST APIs work

API uses the X-Api-Key request header for authentication. After a file or media address is submitted, a task identifier is returned, and the client then polls the corresponding result interface; accounts with specific authorization can also use Webhooks to receive notifications once the task is completed.

1. Create or reset the API Key in the account menu, and store the key only on the server side or in a secure key management system.
2. Depending on the media type, use the appropriate interface for submitting images, videos, or audio, along with the business reference number and any necessary analysis options.
3. Save the identifier of the returned task and poll via the result interface; when a pending status is encountered, use a backoff interval to avoid triggering read restrictions.
4. Analyze the model results and heuristic judgments, and record the model version, parameters, time, as well as the conclusions of manual review for audit purposes.
5. Monitor Credit, failure rate, and duplicate submissions; forcing the cache to be closed will result in the same file being processed again, which may incur additional costs.

Input, output, and file restrictions

The descriptions of the main fields for video and audio interfaces, along with some example of errors, indicate different values regarding the maximum file size; the actual limit should be determined based on the settings of the current account and the results of interface verification. For large files or MCP scenarios, it is preferable to use media addresses, as this helps reduce the size increase caused by Base64 encoding.

Quality requirements for submitted materials

• The image should be as close as possible to the original file, avoiding screen captures, screenshots, further compression, filters from social media platforms, and excessive use of overlapping text.
• Image models are not currently suitable for reliably identifying printed documents, IDs, or receipts; specialized alteration procedures and other verification tools are required for such tasks.
• The voice should be a clear human voice; music, loud traffic noise, and very short audio segments should be avoided. Different speakers and recording devices can also affect the quality of the audio.
• Video frame extraction involves only the analysis of sample frames, so fake segments that appear for a short time may not be captured; for high-risk applications, it is necessary to increase the number of frames and to employ strategies involving audio tracks and key frames.

Credit and rate limits

• Each image or each video frame that is analyzed typically costs 1 Credit; additional requests such as those related to faces, alterations, or heatmaps can increase the cost per item.
• On the dashboard, users can view their account balance, total consumption, and the option to top up funds; however, the specific price per credit unit and the benefits included in each package are not disclosed.
• By default, the read interface is limited to 2 requests per second and 60 requests per minute; the submit interface is limited to 1 request per second and 30 requests per minute.
• Exceeding the limits will result in a 429 error; the client should use exponential backoff. For higher throughput or custom limits, it is necessary to contact sales for confirmation.

Price and subscription

Paid services are handled by third-party marketplaces, and the terms regarding refunds, renewals, and payments may vary depending on the purchase channel and the corporate agreement in place. Individuals subject to EU consumer regulations can exercise their right to withdrawal within 14 days after accepting the terms; however, once the service has been started, it is necessary to familiarize oneself with the terms of the relevant channel as well as the rules regarding the amount already consumed.

Remote MCP Server

identifAI offers remote servers for MCP-compatible agents, encapsulating tasks such as image, video, and audio classification as well as Credit inquiries into 17 different tools. Submitting tasks remains an asynchronous process, and the agent can continue to make requests using the task identifiers returned.

• 6 image tools: file or URL submission, individual and batch results, heatmaps, and manual overlay.
• 5 video tools: submission of files or addresses, individual and batch results, as well as manual override.
• 5 audio tools: submission of files or addresses, individual and batch results, as well as manual override.
• 1 user tool: to check the remaining credit.
• The API Key is passed as a parameter for each tool invocation, and the remote MCP service states that it does not store or cache such keys; in production environments, it is still advisable to avoid including keys directly in ordinary conversations.

Enterprise Agent and Deployment

• The conference agent analyzes live video and audio to provide additional identity risk signals for board meetings, remote authorization processes, customer communications, and government meetings.
• The unified portal is used for configuring detections, real-time monitoring, and viewing reports; however, it is still necessary for the organization to establish procedures for handling alert confirmation, escalation, and interruption meetings.
• Built on Kubernetes, Atlas supports cloud, on-premises enterprise, and hybrid environments, with an emphasis on automated operations, modular expansion, and environmental portability.
• Local deployment compatibility does not mean that it is included by default in all packages; capacity, model updates, hardware, support, and data limits must be specified in the service agreement.

Privacy, zero retention, and data storage

• The product regards the retention of no data as an enterprise capability, and indicates that detection can be carried out without the need to store the original media for an extended period.
• The API offers functions such as duplicate file caching, optional file saving, historical analysis, sample downloading, and result replacement based on file hashes; therefore, the exact scope of zero retention depends on account, endpoint, and tenant configurations.
• The privacy policy covers account information, contact details, professional data, platform usage, system logs, maintenance activities, and device data; accounts are disabled within two days after the end of the business agreement, and they are usually deleted or anonymized within the following 30 days.
• Contracts, invoices, bank records, and business communications may be retained for up to 10 years as required by law, while disputes, security considerations, and legal obligations also lead to an extended period of retention.
• Personal data can be processed by hosting, email, cloud, and professional service providers; the policy states that it will not be transferred outside the European Economic Area or the EU.

The terms also grant identifAI an exclusive, global, and free license to use the outputs for improving the services and ensuring the security of the platform. Before processing sensitive identity data, it is necessary to specify in the enterprise agreement, DPA, and technical configurations what constitutes the original media, caches, outputs, model improvements, logs, and the timing of deletion; reliance solely on the concept of \"no retention\" is not sufficient.

Safety, rights, and responsibilities

• IdentifAI Labs S.P.A. holds ISO 27001 certification, and it applies GDPR, access control, and secure design principles in its corporate deployments.
• Users must have the right to upload content and must comply with privacy, portrait, copyright, and data protection regulations; media detection does not relieve the service provider of its legal responsibilities.
• The user retains the right to input content and owns the output, but the results may be incorrect, incomplete, or not meet expectations; it is up to the user to assess their accuracy and suitability for specific purposes.
• Reverse engineering, crawling, bypassing security measures or usage limits, creating fake accounts, and any other unauthorized use are prohibited.

Open source and client status

IdentifAI’s platform, detection algorithms, models, and conference agents are not fully open-source products; the terms in place prohibit the copying, modification, or disclosure of such software and algorithms. The official GitHub repository contains only the company’s homepage, and no verifiable product SDKs, model weights, or MCP Server licenses available under an open-source license can be found there.

The confirmed channels to date are web pages, REST APIs, remote MCP Servers, enterprise agents, and deployment solutions. Official iOS, Android, or Chrome Store clients released by IdentifAI Labs S.P.A. have not been confirmed yet.

Suitable for users and typical use cases

• KYC and identity team: They detect synthetic selfies, face swapping, and camera injection, taking into account factors such as liveness checks, document verification, and account risk indicators.
• Bank and enterprise security: Detecting voice cloning or identity spoofing in high-risk calls, video conferences, and payment instructions.
• Insurance and claims processing: Analyze the photos, videos, and recordings uploaded by customers; however, special tools as well as manual verification are required to detect any manipulation of these files.
• Media and platforms: Conduct batch screening of suspicious content before publishing or distributing it, and incorporate appeals regarding false positives as well as editorial decisions into the process.
• Agent developers: They can quickly integrate multi-modal detection using MCP, or obtain more detailed parameters, callbacks, and audit controls through REST APIs.

Frequently Asked Questions

Can identifAI determine with 100% certainty whether the content is authentic?

It is not possible; the result is the model’s assessment of the risk associated with human or artificial categories, and this assessment can be affected by factors such as the new generator used, compression, noise, and the quality of the samples. Decisions regarding high risk must be made in conjunction with other evidence and through manual review.

Which media formats is identifAI capable of supporting?

Currently, it supports images, videos, and recordings with voice content; it also offers real-time video conference detection. Text detection is not among the core features available at present.

Is there a free version of identifAI?

The terms allow certain features to be used free of charge, but there are no clear specifications regarding the amount of free credits available, their validity period, or the range of models that can be used. After registration, one should check the account balance; companies that need to use these services must contact sales.

How does an API return results?

After submitting the file or address, the task identifier is returned first, after which polling can be performed via the result interface; Webhooks can also be used once authorization is obtained. The response includes the scores for each model as well as a heuristic-based overall assessment.

Will the uploaded files be saved permanently?

The product claims to have zero data retention, but the platform still offers caching, optional saving, and history functions. The actual behavior depends on the account terms and service agreement; for sensitive materials, it is necessary to check the specific settings and deletion rules first.

Is identifAI open source?

The core platform, algorithms, and models are not open-source products that have been officially recognized. The official GitHub repository does not provide the source code or SDKs for these products, nor does the remote MCP Server offer any verifiable licenses.

Summary

The value of identifAI lies in its ability to integrate visual, audio, and real-time deepfake detection functions within a single corporate platform, while also offering web interfaces, REST APIs, MCP, as well as various deployment options. Before purchasing or implementing it, it is necessary to evaluate false positives and negatives using one’s own samples, to determine the credit limits, throughput capacity, zero-retention policies, and contractual responsibilities, and to ensure that there is a mechanism for manual review.