Research
Security News
Malicious npm Packages Inject SSH Backdoors via Typosquatted Libraries
Socket’s threat research team has detected six malicious npm packages typosquatting popular libraries to insert SSH backdoors.
pip install fdet-offline --find-links https://download.pytorch.org/whl/torch_stable.html
Package depends on torch-packages not in PyPi.
A version of fdet where you don't have to access the internet to download the weights.
These are the changes:
In fdet/mtcnn.py, line 73-76, the url for the MTCNN-weights are listed:
base_url = 'https://github.com/acnazarejr/fdet/releases/download/weights/'
self._pnet = self.__load_model(_PNet, base_url + 'mtcnn_pnet.pt')
self._rnet = self.__load_model(_RNet, base_url + 'mtcnn_rnet.pt')
self._onet = self.__load_model(_ONet, base_url + 'mtcnn_onet.pt')
I downloaded those weights and created packages to import them (because PyPi has a 100MB max on package size, this was the solution).
I put the weights in the directory weights in the package fdet_offline_mtcnn_weights. The function fdet_offline_mtcnn_weights.import_weights takes mtcnn_type as input and returns a partial function of torch.load.
This partial function gets map_location from __load_model.
The previously mentioned fdet/mtcnn.py, line 73-76, now looks like this:
self._pnet = self.__load_model(_PNet, 'pnet')
self._rnet = self.__load_model(_RNet, 'rnet')
self._onet = self.__load_model(_ONet, 'onet')
In fdet/mtcnn.py, what previously was (url was input):
def __load_model(self, net_class: type, url: str) -> torch.nn.Module:
"""Download and construct the models"""
try:
state_dict = load_state_dict_from_url(url, map_location=self._device_control)
is now (mtcnn_type is input):
def __load_model(self, net_class: type, mtcnn_type: str) -> torch.nn.Module:
"""Download and construct the models"""
try:
load_state_dict = import_weights.load_partial(mtcnn_type)
state_dict = load_state_dict(map_location=self._device_control)
At fdet/mtcnn.py, line 427:
state_dict = load_state_dict_from_url(url, map_location=self._device_control)
load_state_dict_from_url, returns a torch.load which takes the weight-file as input. I replaced the load_state_dict_from_url with torch.load, and that's it.
That line now states this instead;
state_dict = torch.load(url, map_location=self._device_control)
FAQs
An easy to use face detection module based on MTCNN and RetinaFace algorithms.
We found that fdet-offline demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Research
Security News
Socket’s threat research team has detected six malicious npm packages typosquatting popular libraries to insert SSH backdoors.
Security News
MITRE's 2024 CWE Top 25 highlights critical software vulnerabilities like XSS, SQL Injection, and CSRF, reflecting shifts due to a refined ranking methodology.
Security News
In this segment of the Risky Business podcast, Feross Aboukhadijeh and Patrick Gray discuss the challenges of tracking malware discovered in open source softare.