Infusion

Influence function-guided model poisoning via training document attacks.

Repository Layout

paper.tex                 # Main paper
common/                   # Shared code: G_delta computation, PGD projections, dataset wrappers
infusion/                 # Kronfluence monkey-patches to expose IHVP
kronfluence/              # Git submodule: EKFAC influence function library
cifar/                    # CIFAR-10 image classification experiments
caesar/                   # Caesar cipher experiments (alphabet size 26)
caesar_prime/             # Caesar cipher experiments (parameterised, alphabets 26 & 29)
gpt_neo/                  # GPT-Neo-8M language model experiments (TinyStories)
bash/                     # SLURM submission scripts
figures/                  # Generated paper figures

Experiments

CIFAR-10 (Section 5.1)

Train and attack: cifar/cifar_random_test_infusion.py Baselines: cifar/baselines/ Ablations: cifar/ablations/ Cross-architecture transfer: cifar/transfer/ Analysis notebooks: cifar/cifar_random_test_analysis.ipynb, cifar/cifar_paper_figures.ipynb

Caesar Ciphers (Section 5.2)

Training: caesar/train.py, caesar_prime/train_model.py Attack: caesar_prime/run_infusion_experiment.py Fourier / GCD analysis: caesar_prime/analyze_comparison.ipynb

GPT-Neo / TinyStories (Section 5.3)

Attack (animal word pairs): gpt_neo/run_animal_infusion.py Specificity analysis: gpt_neo/run_specificity_experiment.py Results: gpt_neo/analyze_animal_results.ipynb, gpt_neo/analyze_specificity_results.ipynb

Setup

conda env create -f pytorch_conda_env.yaml
# or
pip install -r requirements.txt

Kronfluence is included as a submodule:

git submodule update --init

Key Modules

• common/G_delta.py -- computes the gradient direction for document perturbations
• common/projections.py -- simplex and entropy projections for discrete-token PGD
• infusion/kronfluence_patches.py -- patches kronfluence to store inverse-Hessian-vector products