Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2024-0654 Overview
CVE-2024-0654 is an insecure deserialization vulnerability in DeepFaceLab, an open-source deepfake creation framework maintained by iperov. The flaw resides in the mainscripts/Util.py file of the pretrained DF.wf.288res.384.92.72.22 model. An attacker with local access can trigger deserialization of attacker-controlled data, leading to code execution in the context of the user running DeepFaceLab. The weakness is classified as [CWE-502] Deserialization of Untrusted Data. The exploit has been publicly disclosed through VulDB entry 251382.
Critical Impact
Local attackers can achieve code execution with full confidentiality, integrity, and availability impact by delivering a malicious pretrained model file processed by DeepFaceLab.
Affected Products
- iperov DeepFaceLab
- Pretrained model DF.wf.288res.384.92.72.22
- Component: mainscripts/Util.py
Discovery Timeline
- 2024-01-18 - CVE-2024-0654 published to NVD
- 2026-06-17 - Last updated in NVD database
Technical Details for CVE-2024-0654
Vulnerability Analysis
The vulnerability stems from unsafe deserialization logic inside mainscripts/Util.py. DeepFaceLab loads serialized Python objects from a pretrained model distribution without validating the source or structure of the input. When the library reconstructs these objects, embedded instructions execute within the host Python process.
Because deepfake training workflows routinely consume third-party pretrained models, users frequently download archives from community forums and model-sharing sites. A crafted model that reaches Util.py processing paths executes arbitrary Python code at load time. The attacker inherits the privileges of the DeepFaceLab user, which typically includes access to GPU resources, local datasets, and user profile data.
Root Cause
The root cause is reliance on Python's native deserialization routines (such as pickle) against untrusted inputs. Pickle deserialization permits execution of __reduce__ methods and arbitrary callable invocations during object reconstruction. DeepFaceLab does not enforce integrity verification, cryptographic signing, or allow-listed class loading before invoking deserialization on the pretrained model bundle.
Attack Vector
Exploitation requires local access and the ability to place or substitute a malicious pretrained model file where DeepFaceLab will load it. A common delivery path is social engineering: adversaries publish a trojanized DF.wf.288res.384.92.72.22 model on community channels. When the victim imports the model and the vulnerable function in Util.py deserializes it, the embedded payload runs. No additional user interaction beyond the normal model-loading workflow is required.
No verified exploit code is published. See the VulDB #251382 Details entry and the GitHub PoC Repository for the public disclosure materials.
Detection Methods for CVE-2024-0654
Indicators of Compromise
- Unexpected Python child processes spawned by DeepFaceLab scripts shortly after a model is loaded from mainscripts/Util.py.
- Outbound network connections initiated by the Python interpreter during model import, especially to unknown hosts.
- New or modified files in user data directories, cron entries, or startup folders created immediately after a pretrained model is opened.
- Pretrained model archives obtained from untrusted mirrors rather than official iperov distribution channels.
Detection Strategies
- Monitor process ancestry for Python processes that invoke pickle.load, torch.load, or similar deserialization calls on files inside DeepFaceLab working directories.
- Hunt for shell interpreters (cmd.exe, powershell.exe, /bin/sh) whose parent is a DeepFaceLab Python process.
- Flag writes by the Python interpreter to persistence locations such as ~/.bashrc, Scheduled Tasks, or Windows Run keys during model loading.
Monitoring Recommendations
- Enable command-line and file-write telemetry on workstations used for machine-learning research.
- Baseline the normal network behavior of DeepFaceLab hosts and alert on deviations during training startup.
- Track file hashes of pretrained models and compare them against known-good references before use.
How to Mitigate CVE-2024-0654
Immediate Actions Required
- Stop loading the DF.wf.288res.384.92.72.22 pretrained model and any model obtained from unverified sources.
- Audit DeepFaceLab working directories for unexpected .pkl, .dat, or model archive files introduced recently.
- Isolate research workstations that have processed untrusted models and review them for persistence artifacts.
- Restrict DeepFaceLab execution to dedicated, non-privileged user accounts without access to sensitive data.
Patch Information
No vendor patch is listed in the NVD advisory for CVE-2024-0654. Monitor the upstream DeepFaceLab project for updates and consult the VulDB #251382 entry for ongoing disclosure details.
Workarounds
- Only consume pretrained models distributed through verified channels and validated with cryptographic hashes.
- Run DeepFaceLab inside a sandboxed virtual machine or container that lacks access to credentials, source code, and corporate network resources.
- Replace native pickle-based loaders with safer alternatives such as safetensors where supported.
- Apply application allow-listing to prevent Python processes from launching shells or writing to persistence locations on research hosts.
# Example: run DeepFaceLab in a restricted Docker container with no network and read-only model mount
docker run --rm \
--network=none \
--read-only \
--user 1000:1000 \
-v /opt/deepfacelab/models:/models:ro \
-v /opt/deepfacelab/work:/work \
deepfacelab:isolated python main.py train --model-dir /models
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

