Awesome

AnonyNoise: Anonymizing Event Data with Smart Noise to Outsmart Re-Identification and Preserve Privacy (WACV25)

All good in this is from Allah swt and all mistakes are from me or Shaytan.

@InProceedings{bendig2025anonynoise,
  author  = {Bendig, Katharina and Schuster, Ren{\'e} and Thiemer, Nicole and Joisten, Karen and Stricker, Didier},
  title   = {AnonyNoise: Anonymizing Event Data with Smart Noise to Outsmart Re-Identification and Preserve Privacy},
  booktitle = {Winter Conference on Applications of Computer Vision (WACV)},
  year    = {2025},
}

Setup

Install the required dependencies using

pip install -r requirements.txt

You can download the pretrained weights for DVS-Gesture here.

Data

This project uses the following datasets for training and evaluation. Please ensure to download the datasets into ./data and structure them as specified. Preprocessing is required for certain datasets as outlined below.

DVS-Gesture

• you can download the dataset via the tonic library

import tonic 
tonic.datasets.DVSGesture(save_to= './data', train = True)
tonic.datasets.DVSGesture(save_to= './data', train = False)

• Expected Structure:

data/DVSGesture/
├── ibmGestureTrain/
│   ├── user24_flurescent/
│       ├── 0.npy
│       ├── 1.npy
├── ibmGestureTest/
│   ├── user01_flurescent/
│       ├── 0.npy
│       ├── 1.npy

SEE

• Download Link
• Download Link 2
• Expected Structure:

data/SEE/event/
├── angry/
│   ├── 1_001_man_24_Master_normal_angry_45_take000/
│       ├── 00000.jpg
│       ├── 00001.jpg
├── disgust/
│   ├── 7_001_man_24_Master_normal_disgust_44_take001/
│       ├── 00000.jpg
│       ├── 00001.jpg

EventReIDv2

• Download Link
• Preprocessing: you can use the following commands for the preprocessing of data (adapted from ReID without Id)

python ./reid_utils/event_partition.py --input_dir ./Event_ReId_v2/
python ./reid_utils/split_train_test.py  --data_dir ./output/Event_ReId_v2/

• Expected Structure:

data/EventReID_v2/
├── test/
│   ├── 002_c1_001.txt
│   ├── 002_c1_006.txt
├── train/
│   ├── 001_c1_001.txt
│   ├── 001_c1_002.txt

Training

Pretrain Networks

The following comand can be used for pre-training the Re-ID and Target network before the pipeline training. Note: The EventReIDv2 does only contain Re-ID labels and no target ground truth.

DATASETPATH=./data/DVSGesture

python preposttrain_networks.py --ename dvsg_pretrain_target \
    --dataset dvsg --network resnet50 --store_weight --dataset_path ${DATASETPATH} \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type target

python preposttrain_networks.py --ename dvsg_pretrain_id \
    --dataset dvsg --network resnet50 --store_weight --dataset_path ${DATASETPATH} \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type id

Pipeline Training

The following command starts the training of the pipeline. Update BEST_TARGET_WEIGHTS and BEST_ID_WEIGHTS with the paths to the weights of the best performing model from the pre-training step.

BEST_TARGET_WEIGHTS=???
BEST_ID_WEIGHTS=???

python train_pipeline.py --ename dvsg_pipeline \
     --dataset dvsg --network resnet50 --store_weight --dataset_path ${DATASETPATH} \
     --batch_size 32 --lr0 5e-4 --lr0_helper 1e-3 --epochs 300 \
     --target_weights ${BEST_TARGET_WEIGHTS} \
     --id_weights ${BEST_ID_WEIGHTS}

Post Training w/o denoise net

In order to post train the ReID and TargetNetwork based on the anonymized events, the following commands can be used. Update AFTERPIPE_TARGET_WEIGHTS,AFTERPIPE_ID_WEIGHTS and AFTERPIPE_ANON_WEIGHTS with the weights of the last epoch of the pipeline training. Since the anonymization network is trained in an adversarial manner only the last epoch includes the optimal results.

AFTERPIPE_TARGET_WEIGHTS=???
AFTERPIPE_ID_WEIGHTS=???
AFTERPIPE_ANON_WEIGHTS=???

python preposttrain_networks.py --ename dvsg_postpipe_nodn_id \
    --dataset dvsg --network resnet50 --store_weight \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type id \
    --prepostmode post \
    --anon_weights ${AFTERPIPE_ANON_WEIGHTS} \
    --class_weights ${AFTERPIPE_ID_WEIGHTS}

python preposttrain_networks.py --ename dvsg_postpipe_nodn_target \
    --dataset dvsg --network resnet50 --store_weight \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type target \
    --prepostmode post \
    --anon_weights ${AFTERPIPE_ANON_WEIGHTS} \
    --class_weights ${AFTERPIPE_TARGET_WEIGHTS}

Post Training with denoise net

The following includes the commands for the post training inlcuding a denoise network.

python preposttrain_networks.py --ename NEW_dvsg_postpipe_wdn_id \
    --dataset dvsg --network resnet50 --store_weight \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type id \
    --prepostmode post --denoisenet \
    --anon_weights ${AFTERPIPE_ANON_WEIGHTS}  \
    --class_weights ${AFTERPIPE_ID_WEIGHTS} 

python preposttrain_networks.py --ename NEW_dvsg_postpipe_wdn_target \
    --dataset dvsg --network resnet50 --store_weight \
    --batch_size 32 --lr0 1e-4 --epochs 200 --anno_type target \
    --prepostmode post --denoisenet \
    --anon_weights ${AFTERPIPE_ANON_WEIGHTS}  \
    --class_weights ${AFTERPIPE_TARGET_WEIGHTS}

Evaluation

You can evaluate any ReId or Target Network using the following command:

ClASS_WEIGHTS=???
python preposttrain_networks.py --ename dvsg_pretrain_target \
    --dataset dvsg --network resnet50 --dataset_path ${DATASETPATH} \
    --anno_type target --val_only --class_weights ${ClASS_WEIGHTS}

Code Acknowledgements

Our projects partially uses code from the following projects:

• ReID without Id (PDF, GIT) for the ClassNet definition and re-identification training