samuelgong/flashe

[arXiv'21] Additively Symmetric Homomorphic Encryption for Cross-Silo Federated Learning

FLASHE

This repository contains scripts and instructions for reproducing the experiments in our arXiv '21 paper FLASHE. FLASHE is integrated as a pluggable module in FATE v1.2.0, an open-source industrial platform crafted by WeBank for cross-silo federated learning. You may want to use commands like git diff to see what have been changed since the fork (i.e., the first commit).

Overview

• Run Experiments and Validate Results
• Repo Structure
• Acknowledgement
• Contact

Run Experiments and Validate Results

The output of the experiment will validate the following major claims in our background part (section 2 in paper) and evaluation part (section 6 in paper):

Cluster Deployment

1. FLASHE outperforms batching versions of the three baselines by 3.2×-15.1× in iteration time and 2.1×-42.4× in network footprint. Compared to plaintext training, FLASHE achieves near-optimality efficiency with overhead ≤6% and 0% in traffic (§6.2) —> Figure 7 and 8.
2. FLASHE brings down at least 13×-63× of the computation overhead and 48× of the communication overhead of general HEs when sparsification is employed (§6.4) —> Figure 10 and 11.
3. FLASHE's mask precomputation effectively eliminates the cryptographic overhead (rendering it < 0.1s) (§6.6) —> Figure 13.

Local Simulation and Projection

1. The enormous performance of general HE schemes like Paillier, FV, and CKKS raises significant efficiency and scalability concerns. Even with the aid of batch encryption, the message inflation factor remains suboptimal (§2.2 and §2.3) —> Table 2.
2. FLASHE exhibits near-optimality when it comes to economic cost (≤5%). Compared with batching versions of the baselines, the savings are significant (up to 73%-94%) (§6.3) —> Figure 9 and Table 4.
3. FLASHE's double masking scheme achieves optimal latency when the client dropout is mild (§6.5) —> Figure 12.

Cluster Deployment

Please go to ./deployment directory and follow the deployment README to run related scripts.

Local Simulation and Projection

Please go to ./simulation directory and follow the simulation README to run related scripts.

Repo Structure

Repo Root
|---- arch                  # (FATE internal)
|---- cluster-deploy        # (FATE internal)
|---- examples              # Example datasets and configurations of FLASHE
|---- deployment            # Procedures for deploying and evaluating FLASHE
|---- fate_flow             # (FATE internal)
|---- federatedml           # (FATE internal)
|---- simulation            # Procedures related to simulation in FLASHE
|---- utils                 # Tools helpful for playing with FATE/FLASHE
|---- eggroll               # (FATE generated)
|---- fate-flow             # (FATE generated)

Notes

Please consider to cite our paper if you use the code or data in your research project.

@inproceedings{FLASHE-arxiv21,
  title={FLASHE: Additively Symmetric Homomorphic Encryption for Cross-Silo Federated Learning},
  author={Zhifeng Jiang and Wei Wang and Yang Liu},
  booktitle={arXiv:2109.00675},
  year={2021}
}

Recent activity

Acknowledgement

Thanks to Chengliang Zhang and Junzhe Xia for their engineering advice regarding FATE. Thanks to Minchen Yu for his suggestions on development automation.

Contact

Zhifeng Jiang ([email protected])

flashe's Issues

请问您怎么在本地运行demo

不好意思，打扰到您了，请问本地怎么运行demo？
我看了readme，在simulation中，什么也没有，是没有上传文件吗

how to use alphas

您好，关于https://github.com/SamuelGong/FLASHE/blob/master/encrypt_test/final_big_table.ipynb 中的代码有一些问题请教。

对梯度quantize时会返回一个值"

a_q, alphas = quantize([[a]], element_bits)

求和之后unquantize时还需要把alphas作为参数传入，

a_u = unquantize(a_d, alphas, element_bits, num_clients)

这个alphas的作用是什么？

about generate data

您好，在我运行您的代码是，当生成数据是负数的时候，最后的结果是错的
e.g. 我只是简单的将生成的random数，变成了负数
’plaintext = np.random.random(plaintext_length) - 2
最后的结果就不正确了。

加密算法部分

打扰到您了，请问加密算法是否采用的是AES算法呢？如果是用的AES，可是AES算法应该不是同态的吧。

issue about “element_bits,int_bits,additional_bits"

您好,作者!
很抱歉又一次打扰到您.
在您的代码中:https://github.com/SamuelGong/FLASHE/blob/master/encrypt_test/final_big_table.ipynb 请问element_bits,int_bits,additional_bits分别代表什么? 对这些位数的设置有要求吗?
以及在https://github.com/SamuelGong/FLASHE/blob/master/federatedml/secureprotol/jzf_quantize.py 中 关于 batch_size = int_bits // element_bits 这个又是怎么去理解呢?
期待您的回复!

issue about "jzf_flashe.py"

您好，作者！

issue 1

请问下面这个function的目的是什么？

FLASHE/federatedml/secureprotol/jzf_flashe.py

Line 12 in d18303b

def chunks_idx(l, n):

issue 2

在下面这个文件中的line 24-26，meaning of the variables

FLASHE/federatedml/secureprotol/jzf_flashe.py

Line 24 in d18303b

len = end - begin

issue 3

请问如果要将您的这个方案应用至deep Learning中，在每个epoch完成后，需要重新设置这两个吗？

flashe.set_iter_index(epoch)
flashe.idx = 0

劳烦，作者回答上述的3个issue。
如果我有表达不清楚的地方，请您指出，谢谢您！

请问您这个加密部分在哪里呢

能否给出不同的明文在同态加密后计算的接口

您好，由于FATE太过庞大，我想在单台计算机上复现您的论文的隐私保护方案。但是在您的final_big_table.ipynb给出的示例代码中的“Addition”步骤计算的是ciphertext_list = [ciphertext] * num_clients，即num_client个ciphertext求和，这里的ciphertext是相同的，而没有给出不同的ciphertext进行求和的接口。请问能否给出不同的明文在同态加密后计算的接口呢，即一个像ckks.sum()那样通用的接口。

如果可能的话麻烦添加以下内容的演示：
存在两个等长的不同向量A,B，A, B中的元素为[-1, 1]之间的小数

1. 利用flashe加密A, B，得到A_e, B_e；
2. 演示您的同态加密方案可以在密文状态下进行的不同的计算，如A_e+B_e, A_e-B_e, A_e dot B_e......；
3. 解密第二步的计算结果，并与明文状态下的计算结果进行比较。

NameError: name 'pickle' is not defined

作者，您好！
在https://github.com/SamuelGong/FLASHE/blob/master/encrypt_test/final_big_table.ipynb中出现了
NameError: name 'pickle' is not defined

我翻阅您的文件，并没有定义这个对象，请问这个变量是什么？
期待您的回复！