> For the complete documentation index, see [llms.txt](https://library.zoom.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://library.zoom.com/ai/deepfake-risk-detection.md). # Deepfake Risk Detection Daily advances in AI technology are making it more difficult to distinguish between genuine individuals or an imposter; this is a phenomenon known as a deepfake. Recent real-world incidents have demonstrated how bad actors exploit deepfake technology to impersonate others, typically for financial gain or to obtain confidential information. Due to rapid advances in AI, deepfakes are poised to become more common, escalating the associated security threats. Zoom is committed to user safety, so it is introducing a proprietary solution to detect and notify users of potential deepfake risks in Zoom Meetings: Deepfake Risk Detection. {% hint style="warning" %} **Important** This feature is scheduled to be released in December 2026 as a paid add-on SKU. Contact your Zoom representative for details. {% endhint %} Keep reading to learn more about the customer experience, real-world examples, security and privacy context, and practical guidance for users and administrators. What Is Deepfake Risk Detection? Zoom has developed Deepfake Risk Detection, a proprietary methodology, to help determine if potential deepfake risks are present in meetings. The solution is native to Zoom Meetings, making it an embedded user experience without requiring external tools. Deepfake Risk Detection is able to: * Assess the likelihood of synthetic audio or video risks * Notify enabled meeting participants when deepfake risks are identified * Display clear messaging of its assessment of a deepfake to allow participants to evaluate and take any action to mitigate the risks Technical Deep Dive: Zoom’s Unique Approach Zoom has created proprietary machine learning models to detect the likelihood of synthetic video and audio, as compared to a human. When Zoom Deepfake Risk Detection is turned on by either the host or enabled participants, the deepfake detection technology analyzes the active speaker tile. ## How Does Zoom Help Detect Deepfakes? ### Video Detection Zoom Deepfake Risk Detection is run directly on the video stream captured at the sender’s device, before it enters the video processing pipeline, using raw client-side frames. This provides a key advantage that avoids the impact of compression, low-bandwidth video encoding/decoding artifacts, and other quality-degrading processing steps that could obscure or remove deepfake artifacts. We also designed a dual-stream face verification architecture. On the receiver side, the video stream is sent to the cloud and compared against the stream obtained from the sender side to verify that both streams are consistent. This prevents an attacker from sending a genuine video stream from the sender side to the deepfake detection server while delivering a deepfake video stream to a Zoom recipient. Temporal synchronization using timestamps helps ensure that the two video streams are properly aligned before comparison. A key efficiency feature within Deepfake Risk Detection is **adaptive detection frequency**. The system starts with high-frequency checks when a participant joins a meeting, then progressively reduces the frequency over time as confidence builds. Detection rates are further adjusted dynamically based on prior verification results — scaling back when confidence is high and suspending detection when repeated failures suggest the video feed is unreliable. A complementary quality assessment layer evaluates noise, brightness, blur, and landmark detection confidence before running deepfake detection, skipping low-quality frames and returning actionable remediation instructions (for example, improve lighting or reduce motion blur) to the participant, rather than wasting compute on unusable input. In a Zoom meeting, Zoom controls the video filters, avatars, and studio effects in the meeting experience. When Deepfake Risk Detection is enabled, these specific features will be disabled for the duration of the meeting in order to more accurately analyze the underlying audio and video streams. {% hint style="info" %} **Note** Most features under Appearance and Virtual Background **are** compatible with Deepfake Risk Detection. Pose Correction is an exception and will be disabled when Deepfake Risk Detection is enabled. {% endhint %} Data elements of the video and audio streams are captured and sent to the cloud, where detection occurs. The video and audio data and resulting scores are deleted after the meeting ends. ### Audio Detection Zoom Audio Deepfake Risk Detection analyzes the audio for each active speaker in real time on the cloud side. Meeting audio passes through Zoom's standard audio processing before reaching the detection service. A few of these standard processes include echo cancellation, noise suppression, and codec encoding. Our detection model is trained on a wide range of real-world audio conditions, including audio processed by Zoom's own pipeline, to help ensure that genuine speech is not mistakenly flagged as fake. Rather than analyze the entire audio stream continuously, the system intelligently extracts **short speech segments** from the active speaker for classification. A quality assessment layer evaluates audio conditions in real time and automatically excludes segments degraded by poor network quality. This helps avoid analysis on unreliable input and returns resources to more productive use. A **speech recognition gate** helps ensure that only actual speech triggers the deepfake analysis; non-speech audio, such as music or background sounds, is automatically excluded. Similar to video detection, Zoom controls the audio filters and studio effects in the meeting experience. When Deepfake Risk Detection is enabled, these specific features are disabled in order to more accurately analyze the underlying audio streams. However, when a participant uses avatar mode, audio deepfake detection continues to operate, so that speaker verification remains active even when video-based analysis is not available. ## Architecture The diagram below shows an overview of Deepfake Risk Detection.
Architecture diagram showing data flow from sender to client to Zoom Cloud.
## User Experience When you’re in a Zoom Meeting and have Zoom Deepfake Risk Detection enabled, here’s what you can expect to experience if a potential risk is flagged. Zoom Deepfake Risk Detection will notify all enabled users of the potential presence of synthetic audio or video by doing the following: * Flag the suspected participant with a **Deepfake risk** icon * Move that participant to the upper-left quadrant of the screen * Display a **Deepfake risk detected** popup with a link to open the **Deepfake detection dashboard**, which provides additional information The following image shows the icon and popup.
Sample image including Deepfake Risk Detection warnings.
The **Deepfake detection dashboard** provides enabled users with additional information, including: * The risk status of **all** meeting participants * Detection risks and confidence levels display for **both** video and audio streams, providing additional context * Clear calls to action, in order to help inform users of potential next steps The following is an example of the deepfake detection dashboard:
{% hint style="info" %} **Note** Zoom Deepfake Risk Detection also notifies users when a video or audio data stream is weak or too inferior to detect a deepfake, or if a participant did not consent. {% endhint %} Use Cases This section includes three common use cases for Deepfake Risk Detection. If you have a scenario you’d like to discuss, contact your Zoom representative. ### Financial Services Consultant #### Overview Bob Smith is a personal consultant for high-net-worth clients at a private bank. He speaks to five or six clients daily, advising them on wealth strategies and financial transactions to manage their personal finances. #### Risks Bob facilitates money transfers on their behalf, which can include opening up new accounts, making large purchases, and paying estate vendors. Bob relies on his judgement, personally knowing his clients, and light authorization checks. #### Example A long-term client requested an ad-hoc meeting with Bob to inform him of an intended purchase of an exotic car for a family member. The client instructed Bob to wire $190K to the owner. Zoom Deepfake Risk Detection alerted Bob to a high risk of deepfake. Even though this type of purchase wasn’t uncommon for the client, Bob used caution and called the client on his cell phone prior to the actual money transfer. The real client had no knowledge of the meeting and Bob avoided transferring funds to a bad actor. ### HR Recruiter #### Overview Sally Ng is a talent recruiter in a mid-level organization, supporting IT hiring for engineering and technical positions. She has eight to ten open positions at any given time and handles the recruitment process end-to-end, from identifying candidates to negotiating offers. #### Risks Sally is cautious during the recruitment process for remote positions. Her organization is increasingly making offers to candidates offshore. The vetting process is intense, usually with five or six additional interviewers and successful document verification before the final offer is completed. Candidates are often onboarded without meeting someone physically. #### Example Halfway through a promising interview process, Sally connected with the candidate on video to inform them of their final interview round. Zoom Deepfake Detection alerted Sally of a heightened risk, including synthetic audio. Sally cross-checked the identity of the candidate against other sources and determined that the candidate was posing as someone else’s online profile presence. Deepfake Risk Detection was able to avert a situation where a bad actor had access to the company’s internal IT infrastructure and environment. Sally prevented possible ransomware attacks, jeopardized client information, and proprietary software theft. ### Finance Manager #### Overview Tom Cox manages a group of Account Payable specialists for a large organization. His team is responsible for paying vendors that provide services for one of the company’s divisions. #### Risks The company has processes in place to ensure that only approved invoices are paid. The level of approvals needed corresponds to the invoice amount. Some vendors provide critical services that keep the company running, such as infrastructure and data services. Late payments could result in customer disruption, damaging the company’s reputation and/or fines. #### Example Tom received an impromptu phone call from the division Vice President of Finance. The VP instructed Tom to make a $1.2M payment to a vendor for a confidential project. The VP called Tom, since Tom’s boss was out of the office. The payment was late and needed to be made today or Tom’s company would incur late penalties. Being suspicious, Tom asked the VP to join a quick Zoom meeting to verify that it was really him. The VP joined the meeting and, to Tom, looked and sounded identical to the real VP, except that Deepfake Risk Detection notified Tom that it was likely synthetic. Tom refused to make the payment and avoided a costly mistake. ## Enablement and Usage Once the add-on is purchased, Account Owners and Administrators can enable Deepfake Risk Detection at the **Group** level for the organization. To turn on the feature, do the following: 1. Log in to the [Zoom web portal](https://www.zoom.com/). 2. Navigate to the Admin section. Click **Account Management**, then click **Account Settings**. 3. In the top navigation bar, click **General**. Go to the **Security** section and turn on **Enable deepfake risk detection**. {% hint style="info" %} **Note** If you want to lock this setting, so that this feature starts automatically in all meetings, click the lock icon next to the feature toggle. If you’d rather give meeting hosts and entitled users the option to turn the feature on an individual meeting basis, leave the lock open. {% endhint %} To turn on Deepfake Risk Detection in a meeting, go to the menu in the bottom toolbar and click the **Detection** icon.
Image of the menu, showing the Detection option in the middle of the second row.
Consent Both hosts and participants can enable Deepfake Risk Detection for their meetings. All meeting participants will be notified that Deepfake Risk Detection is enabled and will be required to consent. Participants will only need to consent on first use and their consent will be stored in User Settings. Users can always change their consent choice in their Settings portal.
Sample Zoom meeting screen displaying a message that "Deepfake risk detection is on."
## Methodology Zoom is dedicated to ensuring our customers have a trust-worthy and safe meeting experience. As with any risk mitigation solution, nothing is perfect. Bad actors are ever-evolving and will continue to test defenses, so it is worth taking active measures to protect your organization. Using a combination of methods to represent real instances, we performed a head-to-head comparison against leading deepfake solutions for both video and audio. Some methods that we used include: **Video** * Traditional graphics-based editing and compositing methods * Early deep-learning methods, such as auto-encoder-based face swapping * GAN-based generation and manipulation methods * Diffusion-based generation/editing methods * Neural rendering approaches, such as NeRF and 3D gaussian splatting **Audio** * Reference-conditioned generative TTS * Spectrogram-vocoder TTS * Neural codec or vocoder resynthesis * Voice conversion or speech-to-speech {% hint style="warning" %} **Important** Always use caution and discretion, regardless of whether deepfake risk detection is enabled. {% endhint %} --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://library.zoom.com/ai/deepfake-risk-detection.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.