Perspective | Legal Protection Path for Algorithm Optimization - Starting from the Open Source Algorithm Protection of DeepSeek
Published:
2025-02-07
The emergence of DeepSeek has provided a feature-rich and user-friendly AI tool for mainland China. Since its launch at the end of January 2025, its popularity has been continuously increasing, consistently topping the global AI software download charts. The DeepSeek algorithm model significantly enhances energy efficiency and effectiveness by optimizing data collection, comparison, retrieval, and analysis pathways. However, with the global in-depth research on DeepSeek comes a surge of hacker attacks aimed at its core algorithm. This indicates that advancements in algorithm technology also bring legal issues such as intellectual property protection, data privacy risks, and technological abuse. This article starts from the legal needs for algorithm protection, explores the legal regulation pathways in conjunction with the characteristics of algorithms, and proposes a comprehensive protection framework centered on intellectual property law, supported by data security law, and supplemented by algorithm transparency. It also envisions balancing technological innovation and rights protection through scenario-based governance and full-chain supervision.
The emergence of DeepSeek has provided a feature-rich and user-friendly AI tool for mainland China. Since its launch at the end of January 2025, its popularity has been continuously increasing, consistently topping the global AI software download charts. The DeepSeek algorithm model significantly improves energy efficiency and effectiveness by optimizing data collection, comparison, retrieval, and analysis paths. However, with the global in-depth research on DeepSeek comes a surge of hacker attacks aimed at its core algorithm. This shows that advancements in algorithm technology also bring legal issues such as intellectual property protection, data privacy risks, and technology misuse. This article starts from the legal needs of algorithm protection, explores its legal regulation paths in conjunction with the characteristics of algorithms, and proposes a comprehensive protection framework centered on intellectual property law, supported by data security law, and supplemented by algorithm transparency. It envisions balancing technological innovation and rights protection through scenario-based governance and full-chain supervision.
The author also writes this article because they once represented a commercial secret case caused by an algorithm in 2003: a professor invented an algorithm for measuring the diameter of microparticles. By using laser irradiation on a liquid containing microparticles and quickly and accurately measuring the diameter based on the data fed back from the laser, this technology was a significant innovation at the time. The old professor established a high-tech enterprise, wrote the algorithm in assembly language, and installed it in the laser equipment he designed. The device could measure and obtain the diameter of microparticles in the liquid in real-time through the feedback signals from the laser. The device was launched and quickly gained market favor due to its efficiency and affordability. However, due to the old professor's negligence in protecting intellectual property, several employees he hired copied all the software (including the algorithm) onto a USB drive and left. The former employees quickly established several competing companies that eroded market share. When the old professor sought to protect his rights, he found that he had neither registered the copyright of the computer software nor kept good backups of software updates. The patent law and patent examination guidelines of 2003 did not provide protection for pure algorithms, and exhausting civil and criminal means could not fully protect him. Fortunately, it was not too late to mend the fence; a series of subsequent intellectual property protection measures were taken to protect future improvements. It is hoped that this article can provide insights for high-tech enterprises focusing on computer software, algorithms, and other innovations regarding intellectual property protection.
1. Technical Characteristics of Algorithm Optimization and Legal Protection Needs
Through the search of existing materials, the improvements of DeepSeek's algorithm are mainly reflected in: First, the efficiency of data processing has been enhanced. By using technologies such as Multi-Head Latent Attention, the data comparison and analysis paths have been optimized, reducing computational resource consumption and achieving the goal of 'higher performance with fewer resources.' Second, full-chain intelligence: from data collection to result output, the algorithm reduces human intervention through self-supervised learning and reinforcement learning, achieving automation in tasks such as contract clause comparison and case rule extraction. Third, innovative applications of synthetic data, using synthetic data generation technology to avoid privacy risks of original data while enhancing model inference capabilities.
The resulting legal issues are: First, whether innovations in algorithm architecture (such as multi-head latent attention mechanisms) can be included in the scope of patent or trade secret protection, and which method should be chosen for protection? Second, data compliance: does the generation of synthetic data need to follow the privacy protection rules of original data? Third, transparency and interpretability: does the automated decision-making process need to meet the traceability requirements of judicial review?
2. Current Status of Algorithm Protection
1. Limitations of Intellectual Property Protection
First, there are obstacles to patent protection: algorithms are often excluded from patent protection due to their 'abstract idea' nature. For example, although DeepSeek's multi-head latent attention mechanism is innovative, it may be difficult to pass patent examination due to the abstract nature of technical implementation details. As mentioned in the earlier part of this article, the professor's measurement algorithm was not recognized as patentable by the patent law and patent examination guidelines at that time. Although the revised examination guidelines in 2006 stipulated that computer programs or algorithms could apply for patents if they run on certain devices to achieve specific functions, there is still a requirement for a carrier to perform certain functions. The demand for patent protection for algorithms is high and requires full disclosure, which is not conducive to patent protection.
Second, the vulnerability of trade secrets: the trend of open-source algorithms (such as DeepSeek's open-weight model) conflicts with trade secret protection, and open-source agreements may weaken technological exclusivity. Additionally, trade secrets require non-publicity, reasonable protective measures, and practicality, and during rights protection, three identifications of non-publicity, identity, and value are needed, which sets a high threshold. Even more disadvantageous for rights holders is the risk of secondary leakage during the rights protection process.
Finally, the lag in computer software copyright protection: registering software copyright can clarify the ownership and protection content of rights holders, which is currently a common protection method. However, due to the rapid iteration of software or algorithms, it is unrealistic to register with each iteration. Furthermore, copyright infringement cases still require identity identification, which is subjective and easily interfered with by changes in programming languages.
2. Data Privacy and Compliance
Currently, although the synthetic data output by various AI software isdesensitized,if the generation process relies on sensitive data, it still needs to comply with the local laws or data rules of software developers and users. The global nature of users brings cross-border data flow compliance issues, and algorithm optimization relies on multi-source data integration, which may trigger the approval requirements for cross-border transmission under the Data Security Law, such as the 'adequacy determination' standard of the EUGDPR.The algorithm itself does not pose legal risks, but the content it outputs carries compliance risks that should be given due attention.
3. Preliminary Exploration of Legal Paths for Algorithm Protection
1. Build a Multi-Level Intellectual Property Protection System
First, the scope of patent protection should be expanded. The U.S. Algorithm Patent Examination Guidelines can be referenced to include algorithm innovations (such as energy efficiency optimization technology) that are 'significantly and specifically effective' within the scope of patentable subject matter. The National Intellectual Property Administration can clarify the scope of algorithm protection in the new version of the examination guidelines to protect technological innovations through patents.
Second, trade secret protection should be strengthened: the Anti-Unfair Competition Law should prohibit the illegal acquisition of unpublished technical details in open-source code, while improving confidentiality agreements and non-compete clauses. Given that such infringement behaviors are often carried out by hackers, local market supervision bureaus should actively preserve evidence of suspected infringement when handling related unfair competition administrative cases, and if necessary, report to the police for criminal investigation to secure relevant evidence and protect innovation.Third, the computer software copyright registration system should be reformed, focusing on the expression of software source code and functional presentation. For iterative software, registration can be done through remarks on the certificate.2. Improve Data Security and Privacy Protection Mechanisms
For AI platforms, it is recommended to classify and manage synthetic data, categorizing risk levels based on data sources and purposes. High-risk synthetic data should fulfill compliance obligations equivalent to those of original data. At the same time, it is suggested to establish a dynamic cross-border transmission regulatory mechanism and create an algorithm-driven data
risk assessment model.
对于AI平台，建议分类管理合成数据，根据数据来源与用途划分风险等级，高风险合成数据需履行与原始数据同等的合规义务。同时应建议动态化跨境传输监管机制，建立算法驱动的数据风险评估模型Real-time monitoring of the security and compliance of cross-border data flows.
In addition, technological innovation is not a safe harbor for all actions. AI output results still occasionally contain omissions or errors, and there may be issues with algorithmic black boxes and accountability. While protecting algorithms, it is required that algorithm developers disclose core logic and sources of training data in high-risk scenarios such as justice and finance, for example, the transparency rules for 'high-risk AI systems' in the EU's Artificial Intelligence Act. Additionally, referring to China's 'Interim Measures for the Management of Generative Artificial Intelligence Services', companies are required to file algorithm models and accept third-party technical audits, establishing a system for algorithm filing and auditing.
DeepSeek's algorithm optimization represents a shift in artificial intelligence technology from 'scale-driven' to 'efficiency-driven', but its legal protection must balance technological innovation and risk prevention, achieving dynamic balance through the interaction of technical standards and legal rules; establishing mutual recognition mechanisms and compliance reviews for cross-border algorithm applications to avoid fragmented regulation that hinders technological development.
Only by building an inclusive and rigorous legal framework can we achieve a win-win situation for algorithmic technological advancement and rights protection, laying the foundation for the rule of law in the era of artificial intelligence.
Key words:
Related News
Zhongcheng Qingtai Jinan Region
Address: Floor 55-57, Jinan China Resources Center, 11111 Jingshi Road, Lixia District, Jinan City, Shandong Province
