As Homomorphic Encryption schemes like CKKS are not capable of performing non-linear functions such as comparison, other PET techniques, including MPC can provide us with a level of security for a desired machine learning application.
Currently, I'm searching for the current related works that combine CKKS and, MPC; particularly 2PC function secret sharing. Every idea would be greatly appreciated!
Dear Dr. Shokofeh Vahidiansadegh ,
The following articles should be at least partially relevant:
MP2ML: A Mixed-Protocol Machine Learning Framework for Private Inference
https://eprint.iacr.org/2020/721.pdf
_____
_____
Practical MPC+FHE with Applications in Secure Multi-Party Neural Network Evaluation (iacr.org)
https://eprint.iacr.org/2020/550.pdf
_____
_____
Introduction to the CKKS encryption scheme
https://inferati.com/blog/fhe-schemes-ckks
_____
_____
CKKS EXPLAINED, PART 3: ENCRYPTION AND DECRYPTION
Posted on September 14th, 2020 under Homomorphic Encryption
https://blog.openmined.org/ckks-explained-part-3-encryption-and-decryption/
Homomorphic Encryption (HE), such the Cheon-Kim-Kim-Song (CKKS) method, and Multi-Party Computation (MPC), specifically 2PC involving function secret sharing, are exciting topics of research. This combo uses HE and MPC to offer safe and private computations, especially for machine learning. Some thoughts and aspects to consider:
Hybrid Secure Machine Learning Frameworks:
Explore frameworks that leverage CKKS for efficient arithmetic operations on encrypted data and 2PC for HE-challenging operations like non-linear functions. Secure machine learning models like neural networks may benefit from this.
Data Aggregation and Analysis Security:
Investigate multi-source data use cases. CKKS can encrypt data in transit, while MPC, especially 2PC, can compute aggregated data without disclosing individual contributions.
Predictive models that protect privacy:
Develop predictive models with CKKS-encrypted model parameters and HE-2PC prediction or inference. This might be used in privacy-sensitive industries like healthcare and banking.
Hybrid System Benchmarking:
The performance, accuracy, and security of hybrid CKKS and MPC systems should be compared to pure HE or MPC implementations Assess computing speed, data throughput, and resource consumption trade-offs.
Secure Multi-Party ML Protocols:
Create protocols that let multiple parties train or use a machine learning model without sharing data. CKKS secures data during transmission or in an untrusted environment, while 2PC secures model training or evaluation.
Adjusting 2PC for Non-Linear ML Models:
Focus on optimising 2PC protocols to handle non-linear machine learning processes like neural network activations, which CKKS can not support.
Increased Security and Lower Costs:
CKKS-2PC processing and communication overhead should be reduced to make it more feasible for real-world applications. Explore ways to protect the combined system from sophisticated attackers.
Domain-Specific Case Studies:
Try this hybrid technique in finance for private trading algorithms, healthcare for confidential prediction models, and smart cities for secure data analytics.
Operating Complex and Dynamic Data Structures:
Discover how these methods manage dynamic data structures like graphs and trees, which are common in advanced machine learning algorithms.
Standardising Protocols and Libraries:
Develop standard protocols and libraries or contribute to open-source projects to implement hybrid cryptographic approaches.
Due to the complexity and novelty of this study area, collaboration with encryption, data privacy, and machine learning experts would be desirable. Due to the quick evolution of HE and MPC, remaining current is essential.