How to train the model with naive differential privacy?

[lurw2000]

I'm trying to train the model on a small test dataset with naive differential privacy. I tried to change something in the configuration, but I got empty output.
Here is the changed configuration

{
    "global_config": {
        "original_data_file": "../traces/simple-network/Switch1_Ethernet1_to_PC1_Ethernet0-correct/raw.pcap",
        "dataset_type": "pcap",
        "n_chunks": 1,
        "dp": true
    },
    "model_manager": {
        "class": "NetShareManager",
        "config": {
            "pretrain_non_dp": false,
            "pretrain_non_dp_reduce_time": null,
            "pretrain_dp": false
        }
    },
    "model": {
        "class": "DoppelGANgerTFModel",
        "config": {
            "batch_size": 1,
            "sample_len": [
                1
            ],
            "iteration": 80000,
            "extra_checkpoint_freq": 4000,
            "epoch_checkpoint_freq": 1000,
            "gen_feature_num_layers": 1,
            "gen_feature_num_units": 100,
            "gen_attribute_num_layers": 1,
            "gen_attribute_num_units": 32,
            "disc_num_layers": 1,
            "disc_num_units": 32,
            "attr_disc_num_layers": 1,
            "attr_disc_num_units": 32,
            "dp_noise_multiplier": 0.2797,
            "dp_l2_norm_clip": 1.0
        }
    },
    "default": "pcap.json"
}

I suspect it is partly because I set pretrain_dp=false. But if pretrain_dp=true, I will be asked to provide a model pretrained with public dataset.

[sniperyyc]

We have not ported the full DP functionality in this "newer" version of the codebase and hopefully, it will be done soon. Apologize for the potential inconvenience :(

We will keep this issue open and keep you posted once we have finished that part. Thanks for your patience!

[sniperyyc]

We recently refactored the code to PyTorch and the support of differential privacy of Wasserstein loss is, unfortunately, a known open issue in Opacus library. We may find a workaround in the future to support this.