In today’s digital age, especially with the rapid growth of artificial intelligence, data has become a vital resource driving advancements across numerous scientific fields. For instance, the second phase of the Human Genome Project (HGP2) has set an ambitious goal to collect multi-omics data from 1% of the global population—equating to 80 million people out of 8 billion (Liu, 2024). Open science initiatives, which advocate for the open sharing of scientific knowledge, data, and methodologies, have the potential to revolutionize research by facilitating data exchange (Commission, 2024; NASA, 2025; UNESCO, 2025). Significant scientific breakthroughs, such as the rapid development of COVID-19 vaccines, have been greatly accelerated by open data sharing and collaborative efforts within the global scientific community (Zastrow, 2020).

Despite the promising benefits of open science, a fully free and transparent system is not always practical (Scheliga and Friesike, 2014). This issue, often referred to as the “Open Science Dilemma,” is analogous to the classic “Prisoner’s Dilemma” in game theory, where individuals must choose between acting for the collective good or prioritizing personal interests. In the realm of open science, researchers face similar trade-offs. Although open data sharing can greatly benefit the scientific community by fostering faster discovery and wider collaboration, individual researchers often hesitate to participate due to concerns over losing intellectual property rights, recognition, or competitive advantage (Scheliga and Friesike, 2014). Additionally, the absence of sufficient incentives discourages participation, especially when maintaining high data quality requires substantial extra effort. This problem is particularly pronounced in the private sector, where significant volumes of valuable data are held but cannot be shared openly due to IP constraints (Micheli, 2022).

In addition, data safety and privacy concerns, particularly for sensitive human-related data, further compound these challenges. Strict data protection laws in regions such as the U.S., Europe, and China—like HIPAA, GDPR and PIPL—create significant legal and ethical barriers to open sharing (Ness and Committee, 2007; Voigt, 2017; Personal Information Protection Law, 2021). According to IBM’s 2024 Cost of a Data Breach Report, the global average total cost of a data breach is $4.88 million (IBM, 2024). In 2023, the genetic testing company 23andMe experienced a major hacker attack, compromising the personal information of nearly 7 million customers, including sensitive genetic data. This incident not only led to widespread public concern but also sparked class-action lawsuits against the company (Tidy, 2023). Another notable breach is the 2018 data breach at MyHeritage, affecting 92 million users (MyHeritage, 2018). These incidents underscore the vulnerability of biological data security. Studies have shown that even anonymized genetic data can be re-identified through cross-referencing with other publicly available data sources (Bonomi et al., 2020; He and Zhou, 2020).

To address these challenges, we propose a ‘Cryptographic Open Science’ framework that leverages advanced Web 3.0-related cryptography technologies such as blockchain, data encryption (particularly Fully Homomorphic Encryption, FHE) and crypto tokens to create robust data protection mechanisms that safeguard security, intellectual property, and privacy while incentivizing participation (Figure 1). Although it may seem counterintuitive, sharing data with cryptographic protections allows for secure, controlled data exchange without compromising privacy, data ownership, or competitive advantage. This balanced approach between openness and security provides a viable path forward for advancing open science, ensuring that data can be shared in a way that maintains trust and encourages collaboration.

Blockchain is a decentralized, distributed ledger technology that allows data to be recorded, shared, and maintained across a network of computers without the need for a central authority (Crosby et al., 2016; Sarmah, 2018). Each block in the blockchain contains a list of transactions or data entries, a timestamp, and a cryptographic link to the previous block, forming a secure and immutable chain. The use of blockchain techniques ensures that once data is recorded, it cannot be altered without the consensus of the network participants (Figure 2). MedRec is an early exemplar that uses a blockchain layer to manage medical data permissions and access events, while FHIRChain targets clinical interoperability by binding access management to HL7 FHIR-based workflows (Azaria et al., 2016; Zhang et al., 2018).

However, blockchain alone is insufficient to protect data due to the inherent nature of digital assets, where duplication costs are virtually zero. Once data are released publicly, it becomes impossible to maintain control over them, even if data ownership is recorded on blockchains. Privacy concerns and restrictive government policies further compound the challenges of sharing data in plain formats. Before modern cryptographic approaches were widely usable, many biomedical consortia adopted federated query architectures that keep data behind institutional firewalls and only return restricted outputs (typically aggregate counts or vetted summaries). Early systems such as SHRINE operationalized cross-site cohort discovery by sending standardized queries to local clinical repositories and returning aggregate results under institutional review board (IRB) constraints (Weber et al., 2009). These systems provide an important baseline: they preserve institutional control and reduce raw-data transfer, but privacy protection largely relies on governance rules, output restrictions, and trust in the hosting institution rather than cryptographic confidentiality guarantees.

Federated Learning is a machine learning paradigm where a global model is trained collaboratively across multiple decentralized devices or servers holding local data samples, without exchanging the data itself (Kairouz and McMahan, 2021; Mammen, 2021). Each participant trains the model on their local data and shares only the model updates (e.g., gradients) with a central server that aggregates them to improve the global model. Federated learning has achieved significant traction in biomedical applications (Crowson et al., 2022; Rieke et al., 2020). Notable large-scale initiatives include the MELLODDY project, where 10 pharmaceutical companies (including Amgen, Bayer, Janssen, Novartis, etc.) collaboratively trained drug discovery models without sharing proprietary data (Heyndrickx et al., 2023). The experiments involved a data set of 2.6+ billion confidential experimental activity data points cross-pharma companies, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. The results demonstrated markedly higher improvements in multi-partner federated models compared to single-partner models. The Federated Tumor Segmentation (FeTS) initiative spanning 30 international healthcare institutions focusing on training deep learning models for brain tumor segmentation (Pati et al., 2022). While FL enables collaborative model training without data pooling, recent work has exposed critical vulnerabilities—gradient leakage attacks can reconstruct training images (Zhu et al., 2019). Another significant challenge is Statistical Heterogeneity. In healthcare, data distributions vary significantly across institutions due to differences in scanner protocols, patient demographics, and labeling practices. This heterogeneity can lead to slow convergence or poor performance of the global model. Additionally, System Heterogeneity—the varying computational power and network bandwidth of client nodes—poses logistical challenges, as the speed of the training round is often bottlenecked by the slowest participant (Barona López and Borja Saltos, 2025; Li et al., 2025; Shah et al., 2025).

Secure Multi-Party Computation (SMPC) is a set of cryptographic protocols that enable multiple parties to jointly compute a function over their inputs while keeping those inputs private (Zhao et al., 2019; Tran et al., 2021). Each party learns only the output of the computation and gains no additional information about the other parties’ inputs. Cho, et al. introduce a framework that enables large-scale genome-wide association studies (GWAS) to be conducted directly on privately held genomic and phenotypic data using secret sharing and secure multiparty computation. By addressing key computational bottlenecks—particularly quality control and population stratification correction—and achieving linear scalability with respect to cohort size, this work demonstrates that rigorous privacy guarantees can be reconciled with the practical demands of modern GWAS (Cho et al., 2018). Differential privacy (DP) offers a mathematically grounded approach to limiting information leakage from released statistics or trained models by injecting calibrated noise. In biomedical settings, DP is most naturally aligned with releasing aggregates, summary statistics, or privacy-preserving model updates. Yet practical deployments must confront domain-specific constraints: biomedical variables are often highly correlated (e.g., linkage disequilibrium in genetics, repeated measures in EHR trajectories), and naïvely applying DP mechanisms can either weaken privacy guarantees under dependency assumptions or degrade utility significantly. Recent work has explicitly studied DP under dependent records (including genomic-style dependencies), emphasizing that privacy/utility tradeoffs depend strongly on the statistical structure of the data and on how privacy budgets are tracked across repeated queries. This motivates hybrid approaches where DP is applied selectively (e.g., only to certain outputs or query types) and coupled with cryptographic controls that reduce exposure in the first place (Almadhoun et al., 2020). Zero Knowledge Proofs (ZKP)s are cryptographic protocols that allow one party (the prover) to prove to another party (the verifier) that a certain statement is true without revealing any additional information beyond the validity of the statement itself (Sun et al., 2021; Berentsen et al., 2023). ZKPs enable verification of knowledge or attributes without disclosing the underlying data. Trusted Execution Environments (TEE)s are hardware solutions that provide secure areas within a processor that isolate code execution and data processing from the rest of the system (Sabt et al., 2015; Jauernig et al., 2020). They protect sensitive computations from being accessed or tampered with by unauthorized parties, even if the operating system or other software is compromised. While TEEs are generally faster, they require trust in the hardware manufacturer and have been shown to be vulnerable to side-channel attacks (e.g., Spectre, Meltdown), which limits their utility for extremely sensitive genomic data compared to pure cryptographic solutions.

FHE is a form of encryption that allows computations to be performed directly on encrypted data (ciphertexts) without needing to decrypt it first (Figure 3). The result of these computations remains encrypted, and when decrypted by the data owner, matches the result of operations performed on the plaintext (Marcolla et al., 2022; Kun, 2024). This enables secure data processing in untrusted environments such as cloud platforms. Recent advancements in Fully Homomorphic Encryption (FHE) have significantly enhanced the potential for secure and privacy-preserving analysis of genomic and clinical data at scale (Blatt et al., 2020; Kim et al., 2020; Ya et al., 2022; Grishin et al., 2019; Grishin et al., 2021). For example, Cheon et al. who developed an FHE-based method for genome-wide association studies (GWAS), and their approach allows chi-square statistics to be computed on encrypted genomic data, achieving accuracy comparable to traditional methods while preserving data privacy (Kim et al., 2020).

Each of these technologies offers a unique approach to addressing data privacy challenges, and their combined use can ensure the highest level of data protection tailored to specific application scenarios. For example, private keys used to decrypt FHE results can be securely stored within a Trusted Execution Environment (TEE), preventing unauthorized access (Behnke, 2023).

The implementation of COS framework also faces several technical and ethical challenges.

From a technical perspective, while FHE technology has made significant strides recently, it is still evolving and requires further optimization to handle large-scale biological data analysis effectively. The primary challenge for FHE is its computational overhead, with operations being several orders of magnitude slower than plaintext computations (Zhang et al., 2022). Addressing this challenge requires an integrative approach that includes advanced scheme and algorithm development, optimized bootstrapping techniques, and hardware acceleration. Selecting the appropriate FHE scheme based on application needs can significantly improve performance (Marcolla et al., 2022). For instance, the Brakerski-Fan-Vercauteren (BFV, 2012) scheme is optimized for integer arithmetic operations. The Torus FHE (TFHE, 2016) scheme focuses on high-speed bootstrapping and is ideal for binary computations, while the Cheon-Kim-Kim-Song (CKKS, 2017) scheme is designed for approximate computations on floating-point numbers, making it particularly useful for machine learning applications. Both academic and industry players are actively developing FHE libraries for practical applications, including IBM’s HELib, Microsoft’s SEAL, PALISADE, and TFHE (Gouert et al., 2023). Hardware acceleration is another critical factor in enhancing FHE efficiency (Zhang et al., 2022). The development of Field-Programmable Gate Arrays (FPGAs) and Application-Specific Integrated Circuits (ASICs) can greatly speed up FHE operations. ASICs, for example, are estimated to accelerate computations by more than 10,000 times compared to single-core CPU (Ahmad A et al., 2023). Additionally, application-specific optimizations, such as pre-computed lookup tables, can reduce computational requirements, and layered security models can be used to categorize data according to privacy levels, optimizing processing based on the sensitivity of the data. Additionally, transitioning existing bioinformatics algorithms to encrypted computing platforms remains a massive engineering challenge, necessitating the redesign of many classic algorithms. This task demands substantial investments in manpower, time, and close collaboration between biologists, computer scientists, and cryptographers. To make Fully Homomorphic Encryption (FHE) more accessible to non-experts, the community can focus on three key areas of development. First, we can create FHE compilers that allow users to write code in high-level languages like Python, which the compiler then automatically maps to encrypted computations. This approach enables developers to leverage FHE without needing specialized cryptographic knowledge. Second, FHE Software Development Kits (SDKs) can be developed to provide user-friendly, low-level APIs that simplify FHE implementation. While this requires developers to have a basic understanding of homomorphic encryption, they are shielded from its most complex aspects. Finally, we can build bioinformatic algorithm libraries specifically designed for FHE, effectively transferring fundamental algorithmic functions into an encrypted environment. Hosting hackathons and programming competitions could further facilitate collaborations between FHE experts and bioinformaticians to develop such libraries.

While the Cryptographic Open Science (COS) framework offers a technical resolution to the Open Science Dilemma, it introduces a complex set of cyber-ethical challenges that must be navigated to avoid exacerbating existing inequities. A primary ethical concern cited in critical literature is the “hyper-financialization” of health data. By tokenizing genomic information, the COS framework risks transforming patients from altruistic research participants into speculative assets. Additionally, establishing and maintaining a global encrypted biological data sharing platform requires significant investment and sustained international cooperation. Balancing the interests of diverse stakeholders and establishing a sustainable operational model are key challenges. Traditional data sharing initiatives like United Kingdom Biobank and China Kadoorie Biobank rely heavily on public funding and donations. However, the introduction of decentralized Web 3.0 technologies, such as blockchain and token-based incentives, could revolutionize the funding model by encouraging participation and unlocking market-driven investment from the public. This approach has the potential to raise funds at levels far exceeding traditional methods. Critics argue that market-driven incentives might skew research priorities toward commercially viable conditions (e.g., those affecting wealthy populations) at the expense of rare diseases or public health crises affecting marginalized groups. Furthermore, the volatility of crypto-assets could introduce instability into scientific funding, where the value of a research project fluctuates with token market speculation rather than scientific merit. Thirdly, a fundamental tension exists between the immutability of blockchain and the “Right to Erasure” (Right to be Forgotten) mandated by GDPR (Article 17) (Regulation, 2016). Strict interpretations of GDPR suggest that even hashed or encrypted personal data on a blockchain might still qualify as “personal data” if it can technically be linked back to an individual. A possible solution is to utilize “crypto-shredding” (destroying decryption keys) as a compliance mechanism. However, if a blockchain contains immutable references to personal data that a patient has revoked consent for, the entire protocol could still face significant regulatory risk, potentially forcing a “hard fork” or abandonment of the network. Lastly, the complexity of Web 3.0 interfaces poses a significant barrier to equitable access. Expecting patients to manage private keys, wallets, and gas fees for “self-sovereign identity” creates a high technical threshold that may exclude elderly, low-income, or less tech-savvy populations. This could lead to biased datasets that over-represent tech-literate demographics, subsequently training AI models that generalize poorly to the broader population. While FHE offers robust privacy protection, large-scale public education and outreach will be necessary to build understanding and trust in these technologies. Additionally, an international effort is required to harmonize legal and ethical standards across countries and regions, creating a flexible global regulatory framework that respects local requirements while upholding the core principles of data sharing. An international governing body would be needed to ensure that global projects adhere to these standards.

The Cryptographic Open Science framework provides a promising approach to tackling the privacy and security challenges associated with biomedical big data sharing. By integrating advanced cryptographic technologies, decentralized mechanisms, and international collaboration, this framework can facilitate true “open science” while safeguarding individual privacy and data sovereignty. This approach will not only accelerate the development of precision medicine and public health initiatives but also establish a solid data foundation for the application of artificial intelligence in the biomedical field and individual-centric fields like metaverse and digital twins (El Saddik et al., 2025; Saad et al., 2023).