This repo stores the official code for our paper: Revealing and Protecting Labels in Distributed Training (arxiv)
The following scripts export gradients computed by keras image recognition backbones and performs RLG to reveal set of labels in the batch. Results are saved into results_svd.json.
cd image_recognition
python export_gradients.py --model ResNet50 --batch_size 10 --imagenet_dir <path_to_imagenet>
python rlg.py --model ResNet50 --batch_size 10
To reveal the transcript(s) from gradients of lingvo's ASR model, follow these steps:
- Clone lingvo (we use the commit #d5aada292071f17967730ca26d8aea7b8f5a31c2)
- Copy files in
speech_recognition/lingvoover lingvo's code. These files contain minor changes to lingvo's source code to make the model twice differentiable (dynamic -> static RNN, replace while_loop with for-loop). - Build lingvo with bazel (see instructions here, also docker)
- Run this script to reconstruct a single utterance
export LINGVO_RECONSTRUCTION_UTTID=<librispeech_utt_id>
export LINGVO_RECONSTRUCTION_TAG=wpm-rand-bow
export OUTPUT_DIR=/home/trungvd/lingvo-outputs/$LINGVO_RECONSTRUCTION_TAG/$LINGVO_RECONSTRUCTION_UTTID
export REC_PARAMS="--model=asr.librispeech.Librispeech960Wpm \
--checkpoint_path <path_to_a_Librispeech960Wpm_checkpoint> \
--model_tag $LINGVO_RECONSTRUCTION_TAG \
--job trainer \
--mode=async --logdir=$OUTPUT_DIR \
--logtostderr \
--saver_max_to_keep=2 \
--run_locally=gpu \
--learning_rate 0.05 \
--learning_rate_min 0.005 \
--learning_rate_decay_steps 4000 \
--num_unchanged_steps 2000 \
--reconstructed_input atten_context \
--target_unit word"
# Export gradients
../bazel-bin/lingvo/trainer $REC_PARAMS --export_gradient
# Infer bag of words
python ../scripts/infer_bow.py --uttid=$LINGVO_RECONSTRUCTION_UTTID --tag=$LINGVO_RECONSTRUCTION_TAG --target_unit $LINGVO_REC_UNIT
# Transcript reconstruction with BoW
../bazel-bin/lingvo/trainer $REC_PARAMS --use_bow
List of utt ids used for our experiments is stored in speech_recognition/data.
If you use this code in your research, please cite our paper:
@inproceedings{dang2021revealing,
author = {Dang, Trung and Thakkar, Om and Ramaswamy, Swaroop and Mathews, Rajiv and Chin, Peter and Beaufays, Françoise},
booktitle = {Advances in Neural Information Processing Systems},
title = {Revealing and Protecting Labels in Distributed Training},
url = {},
year = {2021}
}