chore(deps): fix 2 critical Dependabot alerts

[shenxingy]

Critical Dependabot alerts remediated

Package	From	To	GHSA	Alert
torch	1.13.1	2.6.0	GHSA-53q9-r3pm-6pq6 (critical, torch.load RCE)	VisionRush#12
pillow	9.5.0	10.2.0	GHSA-3f63-hfp8-52jq (critical)	#3

Companion pin: torchvision 0.14.1 → 0.21.0 (the release paired with torch 2.6.0; it hard-pins torch==2.6.0).

Dockerfile change (required for the bump to be deployable)

torch 2.6.0 requires Python >= 3.9, but the image installed Python 3.8. Switched to python3.10 (newest version compatible with the existing numpy==1.22.0 pin), added python3.10-distutils, fixed the pre-existing update-alternatives typo (/usr/bin/pytho3) so /usr/bin/python3 actually points at 3.10, and switched installs to python3 -m pip so pip targets the right interpreter. Base CUDA image untouched — PyPI torch wheels bundle their own CUDA runtime (2.6.0 defaults to cu124; host driver must support CUDA 12.4, i.e. driver >= 550).

Incidentally cleared (same package bumps)

• torch chore(deps): fix 2 critical Dependabot alerts #7, 可以分享一下训练后权重吗，感谢 VisionRush/DeepFakeDefenders#8 (high, GHSA-pg7h-5qx3-wjr3 / GHSA-5pcm-hx3q-hm94, patched 2.2.0)
• pillow Bump pillow from 9.5.0 to 12.2.0 #1 (high, GHSA-j7hp-h8jx-5ppr, patched 10.0.1), Bump torch from 1.13.1 to 2.8.0 #2 (high, GHSA-8ghj-p4vj-mr35, patched 10.0.0)

Still open, out of scope (critical-only pass)

• pillow Bump pydantic from 1.10.9 to 1.10.13 #4 (high) needs 10.3.0; AttributeError: 'augment_inputs_network' object has no attribute 'module'. Did you mean: 'modules'? VisionRush/DeepFakeDefenders#13/这里另一个参赛选手开源了模型，可以直接使用 VisionRush/DeepFakeDefenders#14 (medium) need 12.2.0
• torch final_model_csv  VisionRush/DeepFakeDefenders#10/训练前发生了问题 VisionRush/DeepFakeDefenders#11/项目需要gcc 8.4? VisionRush/DeepFakeDefenders#18/right train env VisionRush/DeepFakeDefenders#20 need 2.7.1+/2.8/2.9.1/2.10; 文档能写详细点么 VisionRush/DeepFakeDefenders#15/final_model_csv/final_model.pth this model was not found, what do I need to do VisionRush/DeepFakeDefenders#16/有没成功训练完成的，分享一下步骤 VisionRush/DeepFakeDefenders#17/提示显存不够 VisionRush/DeepFakeDefenders#19 have no patched version
• opencv-python update README and fix the replknet training issue VisionRush/DeepFakeDefenders#9 (high), scikit-learn Bump torch from 1.13.1 to 2.12.0 #6, pydantic Bump opencv-python from 4.8.0.74 to 4.8.1.78 #5 (medium)

Behavior note for reviewers

torch 2.6 flips the torch.load default to weights_only=True — this IS the CVE fix. Call sites (main_infer.py:17, merge.py:11-12, toolkit/chelper.py:61,71) load state_dict-style checkpoints, which is the supported pattern; if any existing checkpoint pickle contains non-tensor objects, loading will raise — re-save the checkpoint or pass weights_only=False for trusted local files. main.sh uses torch.distributed.launch --use_env, which is deprecated but still functional in 2.6.

Verification

• uv pip compile requirements.txt --python-version 3.10 resolves cleanly: torch==2.6.0, torchvision==0.21.0, pillow==10.2.0, numpy==1.22.0 all coexist.
• torch 2.6.0, torchvision 0.21.0, pillow 10.2.0 confirmed on PyPI (no heavy ML packages installed locally, per policy).