ferasbg / crypton Goto Github PK

Use trials.sh to write into files the plot data you are getting on the client-wise accuracy/losses, as well as the server-side accuracy/losses. Format the data as such. Format in terms of all those "configurations" you specified in nsl.sh, and so on.

We need to re-write the SGD optimizer used at the client-level, thus modifying the Client objects.

from tensorflow_privacy.privacy.optimizers.dp_optimizer_keras_vectorized import (
    VectorizedDPKerasSGDOptimizer,
)

VectorizedDPKerasSGDOptimizer(
                l2_norm_clip=args.l2_norm_clip,
                noise_multiplier=args.noise_multiplier,
                num_microbatches=args.microbatches,
                learning_rate=args.learning_rate,
)

Purpose

The main purpose is that federated learning alone loosely guarantees confidentiality, and not privacy. Differential privacy is meant to help preserve the privacy of specific details that reveal the source of the data processed on the client-side despite the system's focus on training locally on the client data rather than aggregating it in a centralized manner. In terms of malicious attackers, federated strategies and the architecture of federated learning means little to no utility if the attacker can reveal or deduce personal information from the data at the client-level. It also makes sure that the client models generalize well such that aggregation of generalizations serve as a tool rather than a hinderance.

Correct errors with AdvRegClient and run 10 iterations with client partitions under a default experimental configuration (fl). Make sure that the server logs computes over the entire process.

Use Kubeflow and Tensorflow Extended to setup the pipeline for asynchronous and distributed training across n set of clients in a set of c local clusters.