Conclusion

In this comprehensive review, we establish a systematic definition of Decentralized AI (DeAI) and meticulously examine the challenges and complexities inherent in achieving complete decentralization. We pioneer an exploration into the unique challenges posed by the advent of large-scale models, shedding light on their implications for DeAI ecosystems.

Our analysis delves into various critical domains:

• Data Privacy Preservation: We scrutinize the risks and challenges confronting data providers, presenting an list of techniques spanning privacy learning, federated learning, and cryptography to mitigate privacy concerns effectively.

• Security Attacks: Through a detailed examination, we dissect the list of potential attacks targeting model training in DeAI and survey existing solutions to fortify defenses against such threats.

• Incentive Mechanisms: We delve into strategies aimed at incentivizing data providers and computing powers to sustain high-quality service within the DeAI network, emphasizing the importance of fair evaluation mechanisms. In addition, we discussed the copyright protection techniques for data providers.

• Verification of Computation: Our analysis encompasses techniques designed to verify computation results from computing powers, crucial for safeguarding against fraudulent activities such as fake result attacks.

• Network Communication: We explore diverse solutions geared towards enhancing the efficiency of network communication within decentralized settings, including computation protocol, topology, compression, and parameter-efficient fine-tuning.

Moreover, we confront the dual nature of features exhibited by large models in DeAI:

• While large models exhibit extraordinary memorization capabilities, they also raise significant privacy concerns, particularly regarding the inadvertent memorization of sensitive information.

• Privacy preservation techniques serve as vital safeguards against privacy breaches, yet they may inadvertently obscure the source of origin and hinder copyright protection efforts.

• Data encryption techniques prevents data leakage during network communication, but they also present challenges in auditing malicious data from other parties.

Many existing solutions in the DeAI landscape may not be entirely feasible in the context of large models, given their larger size and enhanced memorization and generalization capabilities. Furthermore, due to the rapid evolution of this field and the vast scope for exploration, some important works may have been overlooked in this survey. For instance, the topic of model encryption warrants further investigation.

To stay updated with the latest version of this survey and to explore emerging topics further, we invite readers to visit our continuously updated repository at https://deai.gitbook.com. We also encourage researchers to collaborate on this open-source GitBook, contributing insightful information to enrich the content.