Abstract
Bitcoin, blockchain, Non-Fungible Token (NFT), Artificial Intelligence (AI) and Chat Generative Pre-Trained Transformer (ChatGPT) have emerged as buzzwords in recent times. We are witnessing the convergence of these evolving technologies to address specific challenges across diverse applications. Blockchain’s decentralization properties have led to a rapid and extensive adoption in key areas like financial services, supply chain, and healthcare. However, in contrast to this decentralized world, AI is largely centralized and typically relies on off-chain data, i.e. data that exists outside the blockchain ecosystem.A significant challenge in the cryptocurrency space revolves around integrating off-chain data and AI with blockchain to improve the accuracy and trustworthiness of cryptocurrency market predictions without losing the decentralization ethos of blockchain. This research envisions a future where the convergence of blockchain technology and predictive AI models creates a novel framework for predicting cryptocurrency prices and AI model verification.
The motivation behind this research stems from the lack of effective prediction and verification methods within the cryptocurrency markets, particularly for leading cryptocurrencies such as Bitcoin and Ethereum. While AI and traditional methods such as technical analysis have shown potential in cryptocurrency price prediction, their effectiveness is limited and tends to decline over time. The decrease in the accuracy of off-chain data sources can be attributed to the complex nature of blockchain systems. In contrast, on-chain data provides a comprehensive and valuable source of information regarding the health of the network, its usage, and the activities of its users. The data provided can be a valuable resource for AI models, allowing them to effectively analyze trends, predict behaviours, and identify bottlenecks with enhanced accuracy. The lack of inherent trust in decentralized settings, coupled with privacy concerns and transparency challenges, makes it difficult to ensure the integrity and accuracy of AI models. Traditional trust mechanisms, which often rely on centralization, conflict with the decentralized ethos of blockchain, creating a challenging environment for trust in AI. This gap highlights the need for a novel framework that predicts cryptocurrency prices accurately and verifies the accuracy of the AI model predictions while preserving the privacy of the AI model.
The key objectives of this thesis are: 1. To improve the accuracy and efficiency of AI models in predicting cryptocurrency prices, primarily Bitcoin and Ethereum.
2. To enhance the trust of AI models on blockchain by verifying the integrity and accuracy of AI model predictions for cryptocurrencies while preserving the privacy of AI models.
These objectives are met by developing a comprehensive framework that utilizes on-chain data to improve the understanding of blockchain dynamics. The most significant on-chain data was used as input for the Long Short Term Memory (LSTM) model for price prediction. A self-adaptive framework was developed to tune its hyperparameters to find the most optimized model efficiently, providing more accurate predictions. Subsequently, a verification framework using privacy-preserving techniques such as Zero-Knowledge proofs (ZKP) is developed and implemented to enable decentralized verification of AI model predictions on blockchain. The contributions of this thesis can be summarized as:
• Investigated the significance of on-chain data using an on-chain analysis approach for permissionless blockchain platforms like Bitcoin, Ethereum and private blockchains.
• Developed a novel self-adaptive framework for improved cryptocurrency price prediction, primarily for Bitcoin and Ethereum, outperforming traditional models in accuracy and efficiency by dynamically tuning its hyperparameters.
• Developed a novel verification framework that leverages decentralized oracles and Zero-Knowledge Succinct Non-Interactive Argument of Knowledge (zk-SNARKs) to verify the performance claims of AI models on blockchain while preserving the privacy of AI models.
• The verification framework was implemented in a real-world system by integrating zk-SNARKs with Chainlink oracles, demonstrating their feasibility inAI model verification in decentralized systems.
• Identified key areas for zk-SNARKs optimization by analyzing the efficiency and resource consumption of zk-SNARKs proof generation and verification in blockchain systems.
• Analyzed the computational costs associated with zk-SNARKs verifications such as transaction fees in blockchain and LINK token costs in Chainlink oracles.
| Date of Award | 2025 |
|---|---|
| Original language | English |
| Supervisor | Kumudu MUNASINGHE (Supervisor) & Dharmendra Sharma AM PhD (Supervisor) |