Abstract:The need for data security and model integrity has been accentuated by the rapid adoption of AI and ML in data-driven domains including healthcare, finance, and security. Large models are crucial for tasks like diagnosing diseases and forecasting finances but tend to be delicate and not very scalable. Decentralized systems solve this issue by distributing the workload and reducing central points of failure. Yet, data and processes spread across different nodes can be at risk of unauthorized access, especially when they involve sensitive information. Nesa solves these challenges with a comprehensive framework using multiple techniques to protect data and model outputs. This includes zero-knowledge proofs for secure model verification. The framework also introduces consensus-based verification checks for consistent outputs across nodes and confirms model integrity. Split Learning divides models into segments processed by different nodes for data privacy by preventing full data access at any single point. For hardware-based security, trusted execution environments are used to protect data and computations within secure zones. Nesa's state-of-the-art proofs and principles demonstrate the framework's effectiveness, making it a promising approach for securely democratizing artificial intelligence.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to ensure data security and model integrity when performing artificial intelligence (AI) inference in decentralized systems. Specifically, the paper focuses on the following aspects: 1. **Data Security and Privacy Protection**: - In a decentralized system, data and processing are distributed across different nodes, which may lead to unauthorized access, especially when sensitive information is involved. - Users need to ensure that their input data will not be directly exposed to the decentralized nodes performing AI inference. 2. **Model Integrity and Security**: - Ensure the integrity and authenticity of inference results and prevent malicious nodes from tampering with or forging inference results. - Service providers need to prove to their customers that the output results are generated by a verified large - language model (such as GPT - 4), rather than relying on human - written or other less advanced models (such as GPT - 3.5). 3. **Efficient Security Mechanisms**: - While ensuring high security and privacy, minimize the impact on inference speed and computational resources, especially when dealing with large - scale data and complex models. To solve these problems, the paper proposes a comprehensive framework that utilizes multiple techniques to protect data and model outputs, including but not limited to: - **Zero - Knowledge Proofs (ZKPs)**: Used for secure model verification, ensuring that each node can prove the correctness of its execution without revealing model details. - **Consensus - Based Verification (CBV)**: Ensures the consistency of inference results across nodes. - **Split Learning**: Divides the model into multiple parts to be processed by different nodes, preventing any single node from obtaining complete data. - **Sequential Homomorphic Encryption (SHE)**: Encrypts operations in neural networks to prevent intermediate representations from being extracted by attackers. - **Trusted Execution Environments (TEEs)**: Provide hardware - level secure isolation zones to protect data and computational processes. The application of these techniques enables the Nesa platform to achieve efficient, secure, and private AI inference in a decentralized environment, suitable for critical areas from medical diagnosis to financial decision - making.

Complete Security and Privacy for AI Inference in Decentralized Systems

A Three-layer Security Assurance Model for a Decentralized Federated Learning System

Privacy-Preserving Collaborative Deep Learning with Unreliable Participants.

Poster: Nebula: an Industrial-purpose Privacy-preserving Machine Learning System

Model Agnostic Hybrid Sharding For Heterogeneous Distributed Inference

Industrial Scale Privacy Preserving Deep Neural Network

SAKSHI: Decentralized AI Platforms

FairProof : Confidential and Certifiable Fairness for Neural Networks

Towards Scalable and Privacy-Preserving Deep Neural Network via Algorithmic-Cryptographic Co-design

SafetyNets: Verifiable Execution of Deep Neural Networks on an Untrusted Cloud

Quantitative Verification of Neural Networks And its Security Applications

Model Protection: Real-Time Privacy-Preserving Inference Service for Model Privacy at the Edge

EdgeSanitizer: Locally Differentially Private Deep Inference at the Edge for Mobile Data Analytics

Toward Scalable and Privacy-preserving Deep Neural Network Via Algorithmic-Cryptographic Co-design

A survey on secure decentralized optimization and learning

Penetralium: Privacy-preserving and memory-efficient neural network inference at the edge

A Secure and Privacy-Preserving Machine Learning Model Sharing Scheme for Edge-Enabled IoT

$S^2$NeRF: Privacy-preserving Training Framework for NeRF

SecureML: A System for Scalable Privacy-Preserving Machine Learning

Decentralized Federated Learning Preserves Model and Data Privacy

Decentralized Machine Learning Framework for the Internet of Things: Enhancing Security, Privacy, and Efficiency in Cloud-Integrated Environments