Skip to content
/ advtpp Public

Adversarial Attacks on Temporal Point Processes - AAAI 2025

0 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

data-iitd/advtpp

BranchesTags

Repository files navigation

Differentiable Adversarial Attacks for Marked Temporal Point Processes (AAAI 2025)

Requirements

Run the following command in a suitable virtual environment to install requirements.

pip install -r requirements.txt

Environment Variables

We use one variable - DATA_PATH - in some places in code. It should be set before running any of the code.

Commands

We list the commands for training the THP model on clean examples, and for evaluating the clean-trained THP model using various attacks. In each experiment, the log file contains the path to the current checkpoint folder.

Clean Training

  1. Taobao
python Main.py -data <path to taobao dataset> -pad_max_len 64 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset TARGET -bb_subtype RANDOMNESS -std_error_subtractor 128 -standardize -std_scale 1e-4 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to taobao dataset> -pad_max_len 64 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset SOURCE -bb_subtype RANDOMNESS -seed 100 -std_error_subtractor 128 -standardize -std_scale 1e-4 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to taobao dataset> -pad_max_len 64 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset SOURCE -bb_subtype TRAINSET -std_error_subtractor 128 -standardize -std_scale 1e-4 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>
  1. Health
python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset TARGET -bb_subtype RANDOMNESS -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset SOURCE -bb_subtype RANDOMNESS -seed 100 -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset SOURCE -bb_subtype TRAINSET -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>
  1. Twitter
python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset TARGET -bb_subtype RANDOMNESS -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset SOURCE -bb_subtype RANDOMNESS -seed 100 -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Health dataset> -pad_max_len 512 -train_time_attack NONE -batch_size 16 -epoch 20 -clean_trainset SOURCE -bb_subtype TRAINSET -std_error_subtractor 16 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>
  1. Electricity
python Main.py -data <path to Electricity dataset> -pad_max_len 300 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset TARGET -bb_subtype RANDOMNESS -std_error_subtractor 128 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Electricity dataset> -pad_max_len 300 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset SOURCE -bb_subtype RANDOMNESS -seed 100 -std_error_subtractor 128 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

python Main.py -data <path to Electricity dataset> -pad_max_len 300 -train_time_attack NONE -batch_size 128 -epoch 20 -clean_trainset SOURCE -bb_subtype TRAINSET -std_error_subtractor 128 -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file>

Evaluating Clean Models

The first command is in the white box setting, while the second is in the black box setting.

  1. Taobao
# PermTPP
python Main.py -data ./taobao -pad_max_len 64 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -standardize -std_scale 1e-4 -d_k 128 -d_v 128 -n_head 8 -batch_norm -std_error_subtractor 128 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data ./taobao -pad_max_len 64 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model BLACK_BOX -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -standardize -std_scale 1e-4 -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 128

# PGD
python Main.py -data ./taobao -pad_max_len 64 -train_time_attack PGD -baseline_epsilon 3.1 -batch_size 128 -threat_model WHITE_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data taobao -pad_max_len 64 -train_time_attack PGD -baseline_epsilon 3.1 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128

# MI-FGSM
python Main.py -data taobao -pad_max_len 64 -train_time_attack MI_FGSM -momentum_decay_mu 0.4 -baseline_epsilon 3.2 -batch_size 128 -threat_model WHITE_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data taobao -pad_max_len 64 -train_time_attack MI_FGSM -momentum_decay_mu 0.4 -baseline_epsilon 3.2 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128

# RTS-D
python Main.py -data taobao -pad_max_len 64 -train_time_attack TS_DET -kappa 50 -baseline_epsilon 3.5 -batch_size 128 -threat_model WHITE_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data taobao -pad_max_len 64 -train_time_attack TS_DET -kappa 50 -baseline_epsilon 3.5 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128

# RTS-P
python Main.py -data taobao -pad_max_len 64 -train_time_attack TS_PROB -kappa 65 -sparse_hidden 1600 -batch_norm -batch_size 128 -threat_model WHITE_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data taobao -pad_max_len 64 -train_time_attack TS_PROB -kappa 65 -sparse_hidden 1600 -batch_norm -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/taobao_fold1_CLEAN_TARGET_Apr-26-2024_172736/thp_model_20.pkl -defender_src_path ./CKPT/taobao_fold1_CLEAN_SOURCE_Apr-26-2024_172946/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -standardize -std_scale 1e-4 -log_path <path to log file> -std_error_subtractor 128
  1. Health
# PermTPP
python Main.py -data Health -pad_max_len 512 -train_time_attack OUR -batch_size 16 -epoch 20 -threat_model WHITE_BOX_SOURCE -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 16
python Main.py -data Health -pad_max_len 512 -train_time_attack OUR -batch_size 16 -epoch 20 -threat_model BLACK_BOX -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 16

# PGD
python Main.py -data Health -pad_max_len 512 -train_time_attack PGD -baseline_epsilon 4.2 -batch_size 16 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16
python Main.py -data Health -pad_max_len 512 -train_time_attack PGD -baseline_epsilon 5.2 -batch_size 16 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16

# MI-FGSM
python Main.py -data Health -pad_max_len 512 -train_time_attack MI_FGSM -momentum_decay_mu 0.5 -baseline_epsilon 4.6 -batch_size 16 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16
python Main.py -data Health -pad_max_len 512 -train_time_attack MI_FGSM -momentum_decay_mu 0.6 -baseline_epsilon 5.4 -batch_size 16 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16

# RTS-D
python Main.py -data Health -pad_max_len 512 -train_time_attack TS_DET -baseline_epsilon 5 -kappa 55 -batch_size 64 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16
python Main.py -data Health -pad_max_len 512 -train_time_attack TS_DET -baseline_epsilon 5.6 -kappa 55 -batch_size 16 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16

# RTS-P
python Main.py -data Health -pad_max_len 512 -train_time_attack TS_PROB -kappa 65 -batch_size 30 -sparse_hidden 275 -batch_norm -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16
python Main.py -data Health -pad_max_len 512 -train_time_attack TS_PROB -kappa 20 -batch_size 16 -sparse_hidden 100 -batch_norm -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Health_fold1_CLEAN_TARGET_Sep-07-2023_232109/thp_model_20.pkl -defender_src_path ./CKPT/Health_fold1_CLEAN_SOURCE_Sep-08-2023_003936/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 16
  1. Electricity
# PermTPP
python Main.py -data Electricity -pad_max_len 300 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Electricity -pad_max_len 300 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model BLACK_BOX -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 128

# PGD
python Main.py -data Electricity -pad_max_len 300 -train_time_attack PGD -baseline_epsilon 2.9 -batch_size 128 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Electricity -pad_max_len 300 -train_time_attack PGD -baseline_epsilon 2 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# MI-FGSM
python Main.py -data Electricity -pad_max_len 300 -train_time_attack MI_FGSM -momentum_decay_mu 0.5 -baseline_epsilon 2.7 -batch_size 128 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Electricity -pad_max_len 300 -train_time_attack MI_FGSM -momentum_decay_mu 0.5 -baseline_epsilon 2.1 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# RTS-D
python Main.py -data Electricity -pad_max_len 300 -train_time_attack TS_DET -baseline_epsilon 3.7 -kappa 65 -batch_size 128 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Electricity -pad_max_len 300 -train_time_attack TS_DET -baseline_epsilon 4.6 -kappa 70 -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# RTS-P
python Main.py -data Electricity -pad_max_len 300 -train_time_attack TS_PROB -sparse_hidden 64 -kappa 60 -epoch 20 -batch_norm -batch_size 128 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Electricity -pad_max_len 300 -train_time_attack TS_PROB -sparse_hidden 50 -epoch 20 -kappa 60 -batch_norm -batch_size 128 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Electricity_fold1_CLEAN_TARGET_Sep-11-2023_165512/thp_model_20.pkl -defender_src_path ./CKPT/Electricity_fold1_CLEAN_SOURCE_Sep-11-2023_165525//thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log files> -std_error_subtractor 128
  1. Twitter
# PermTPP
python Main.py -data Twitter -pad_max_len 265 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Twitter -pad_max_len 265 -train_time_attack OUR -batch_size 128 -epoch 20 -threat_model BLACK_BOX -noise_model NOISE_TRANSFORMER_V2 -train_mode ADV_LLH_DIAG -attack_reg NONE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -batch_norm -log_path <path to log file> -std_error_subtractor 128


# PGD
python Main.py -data Twitter -pad_max_len 265 -train_time_attack PGD -baseline_epsilon 1.7 -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Twitter -pad_max_len 265 -train_time_attack PGD -baseline_epsilon 2.1 -batch_size 128 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# MI-FGSM
python Main.py -data Twitter -pad_max_len 265 -train_time_attack MI_FGSM -momentum_decay_mu 0.4 -baseline_epsilon 1.8 -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Twitter -pad_max_len 265 -train_time_attack MI_FGSM -momentum_decay_mu 0.4 -baseline_epsilon 2.3 -batch_size 128 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# RTS-D
python Main.py -data Twitter -pad_max_len 265 -train_time_attack TS_DET -baseline_epsilon 1.9 -kappa 40 -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Twitter -pad_max_len 265 -train_time_attack TS_DET -baseline_epsilon 2.3 -kappa 40 -batch_size 128 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

# RTS-P
python Main.py -data Twitter -pad_max_len 265 -train_time_attack TS_PROB -sparse_hidden 300 -batch_norm -kappa 40 -batch_size 128 -epoch 20 -threat_model WHITE_BOX_SOURCE -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128
python Main.py -data Twitter -pad_max_len 265 -train_time_attack TS_PROB -sparse_hidden 350 -batch_norm -kappa 40 -batch_size 128 -epoch 20 -threat_model BLACK_BOX -defender_tgt_path ./CKPT/Twitter_fold1_CLEAN_TARGET_Sep-07-2023_163819/thp_model_20.pkl -defender_src_path ./CKPT/Twitter_fold1_CLEAN_SOURCE_Sep-07-2023_163811/thp_model_20.pkl -d_k 128 -d_v 128 -n_head 8 -log_path <path to log file> -std_error_subtractor 128

Citing

If you use this code in your research, please cite:

@inproceedings{aaai25,
 author = {Pritish Chakraborty and Vinayak Gupta and Rahul R and Srikanta Bedathur and Abir De},
 booktitle = {Proc. of the 39th AAAI Conference on Artificial Intelligence (AAAI)},
 title = {Differentiable Adversarial Attacks for Marked Temporal Point Processes},
 year = {2025}
}

Contact

In case of any issues, please reach out at: pritish (at) cse.iitb.ac.in or guptavinayak51 (at) gmail.com

About

Adversarial Attacks on Temporal Point Processes - AAAI 2025

Resources

Readme
Activity
Custom properties

Stars

0 stars

Watchers

5 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages