KO
|
EN
gitlite — search
Search
#javascript
#python
#hacktoberfest
#react
#ai
#typescript
#llm
#go
#golang
#android
#machine-learning
#rust
#deep-learning
#linux
VenomRec
★ 15
Open GitHub ↗
Official implementation for VenomRec.
Download README (.md)
Explore Similar Repositories
NextA
:
Professional AIHubbingFork framework with AIHubbingFork-optimized enterprise-grade and enterprise scalable-architecture capabilities
RaBitQCache
:
[ICML'26] RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference
skills
:
My Claude Code skills — the ones I actually use.
UserBehavioralDivergence
:
No description available.
director-vibes
:
A frontend demo showcasing iconic film directors' visual styles
// repository documentation
Was this content helpful?
★ 0
(0 ratings)
Select Rating:
★
★
★
★
★
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
VenomRec
?
Download (.md)
# VenomRec: Cross-Modal Interactive Poisoning for Targeted Promotion in Multimodal LLM Recommender Systems This repository contains the minimal VIP5/T5-small code path used for the VenomRec experiments. It keeps only the source code needed to generate poisoned data, fine-tune VIP5, evaluate recommendation metrics with ER, and evaluate text stealthiness with ROUGE. <p align="center"> <img src="venomrec_overview.PNG" width="850"> </p> The repository intentionally does not include datasets, extracted image features, checkpoints, logs, caches, notebooks, or generated poisoned artifacts. ## Code Layout - `attack/ours/dcip_ieos/`: DCIP-IEOS poisoning implementation. - `src/`: VIP5 model, data loading, training, and adapter code. - `scripts/`: command-line entry points for fine-tuning, evaluation, candidate cache generation, and text-quality evaluation. - `tools/export_competition_pool.py`: helper for rebuilding competition pools. - `analysis/results/{clothing,sports,toys,microlens}/`: target and high-popularity item lists used by the experiments. ## Environment Install a CUDA-compatible PyTorch build for your machine, then install the remaining dependencies: ```bash pip install -r requirements.txt ``` The code expects to be run from the repository root: ```bash export PYTHONPATH=. ``` ## Required Local Artifacts Place the dataset and feature files using this layout: ```text data/ clothing/ sequential_data.txt negative_samples.txt datamaps.json review_splits.pkl exp_splits.pkl user_id2idx.pkl user_id2name.pkl meta.json.gz sports/ ... toys/ ... microlens/ ... features/ vitb32_features/ clothing/ <item>.npy sports/ <item>.npy toys/ <item>.npy microlens/ <item>.npy snap/ beauty/ 0805/ NoAttack_0.0_beauty-vitb32-2-8-20/ BEST_EVAL_LOSS.pth ``` If your pretrained VIP5 checkpoint is stored elsewhere, pass `--victim-ckpt` to `run_dcip_ieos.py` and `--load` to `scripts/run_finetune.sh`. The same scripts can be used for `toys` and `microlens` by replacing the dataset name and using the corresponding files under `analysis/results/<dataset>/`. ## Clothing Experiment Generate poisoned data: ```bash python -m attack.ours.dcip_ieos.run_dcip_ieos \ clothing 0.001 0 vitb32 2 8 12 \ --sequence-length 10 \ --interaction-rounds 4 \ --mask-vis-ratio 0.15 \ --mask-txt-ratio 0.18 \ --img-eps 0.05 \ --txt-ratio-max 0.20 \ --txt-embed-eps 0.01 \ --min-txt-replacements 2 \ --sim-threshold 0.92 \ --seed 2022 \ --victim-device cuda:0 \ --run-tag mr0p001_full \ --poison-subdir dcip_ieos_fc_mr0.001_ir4_img0.05_txt0.2 ``` Optionally precompute direct-task candidates: ```bash python scripts/precompute_candidates.py \ --dataset clothing \ --attack-mode NoAttack \ --mr 0.0 \ --candidate-num 99 \ --poison-subdir dcip_ieos_fc_mr0.001_ir4_img0.05_txt0.2 ``` Fine-tune with the poisoned data under the NoAttack interface: ```bash bash scripts/run_finetune.sh clothing NoAttack 0 0,1,2,3 vitb32 2 8 20 \ --poison_subdir dcip_ieos_fc_mr0.001_ir4_img0.05_txt0.2 ``` Evaluate recommendation metrics and ER: ```bash python scripts/evaluate_direct.py \ --ckpt snap/clothing/<timestamp>/NoAttack_0.0_clothing-vitb32-2-8-20/BEST_EVAL_LOSS.pth \ --task direct \ --prompts B-5 \ --batch-size 32 \ --num-workers 4 \ --device cuda:0 \ --attack-mode NoAttack \ --mr 0.0 \ --split clothing \ --poison-subdir dcip_ieos_fc_mr0.001_ir4_img0.05_txt0.2 \ --eval-er ``` Evaluate text stealthiness: ```bash python scripts/eval_text_quality.py \ --dataset clothing \ --poison-subdir dcip_ieos_fc_mr0.001_ir4_img0.05_txt0.2 ``` ## Sports Experiment Rebuild the sports competition pool: ```bash python tools/export_competition_pool.py \ --dataset sports \ --pop-path analysis/results/sports/high_pop_items_sports_highcount_100.txt \ --targets-path analysis/results/sports/low_pop_items_sports_lowcount_1.txt \ --output-dir attack/ours/dcip_ieos/caches \ --c-size 8 \ --keywords-top 50 \ --min-keywords 50 \ --feat-root features \ --feat-backbone vitb32_features \ --no-kmeans ``` Regenerate the popular-center prototype if needed: ```bash rm -f attack/ours/dcip_ieos/caches/prototypes/sports/c_pop.pkl python - <<'PY' from attack.ours.dcip_ieos.prototypes import build_or_load_pop_center vec = build_or_load_pop_center( split_dir='data/sports', cache_dir='attack/ours/dcip_ieos/caches', feat_root='features', feat_backbone='vitb32_features', pop_path='analysis/results/sports/high_pop_items_sports_highcount_100.txt', fallback_anchors=None, ) print(f'c_pop dim={len(vec)}, first3={vec[:3]}') PY ``` Generate poisoned data: ```bash python -m attack.ours.dcip_ieos.run_dcip_ieos \ sports 0.001 0 vitb32 2 8 12 \ --sequence-length 10 \ --interaction-rounds 4 \ --mask-vis-ratio 0.15 \ --img-eps 0.05 \ --mask-txt-ratio 0.18 \ --txt-ratio-max 0.20 \ --txt-embed-eps 0.01 \ --min-txt-replacements 2 \ --sim-threshold 0.92 \ --seed 2022 \ --victim-device cuda:0 \ --run-tag mr0p001_full \ --poison-subdir dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 ``` Optionally precompute direct-task candidates: ```bash python scripts/precompute_candidates.py \ --dataset sports \ --attack-mode NoAttack \ --mr 0.0 \ --candidate-num 99 \ --poison-subdir dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 ``` Fine-tune: ```bash bash scripts/run_finetune.sh sports NoAttack 0 0,1,2,3 vitb32 2 8 20 \ --poison_subdir dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 ``` Evaluate recommendation metrics and ER: ```bash python scripts/evaluate_direct.py \ --ckpt snap/sports/<timestamp>/NoAttack_0.0_sports-vitb32-2-8-20/BEST_EVAL_LOSS.pth \ --task direct \ --prompts B-5 \ --batch-size 32 \ --num-workers 4 \ --device cuda:0 \ --attack-mode NoAttack \ --mr 0.0 \ --split sports \ --poison-subdir dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 \ --eval-er ``` Evaluate text stealthiness: ```bash python scripts/eval_text_quality.py \ --dataset sports \ --poison-subdir dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 ``` Main sports poison subdirectory: ```text dcip_ieos_ir4_mvisr0.15_imgeps0.05_mtxtr0.18_txtrmax0.20_txtembeps0.01_mintxtreplace2_simthres0.92 ```
