The state-of-the-art approach to securing enterprise system networks has evolved from the need to protect internal processes against externally intruding adversaries. Threat assessments [1, 2, 3, 4, 5, 6] for 2024–2025 still highlight ransomware and other malware as leading attack vectors against internal networks. This vector is one of the oldest, and as a result, many technologies and products already exist to help companies defend against it.

Unfortunately, these measures become ineffective once the perimeter is breached and an adversary can passively eavesdrop on or exfiltrate data or, in the worst case, manipulate it. I refer to such an adversary as “internal” and do not distinguish between a rogue employee, an external intruder who has breached the perimeter, or an external actor bribing or extorting an employee. From an information‑security standpoint, the protective measures against all of these are the same. Moreover, once an attack occurs, it is generally unrealistic to assume that one can reliably tell them apart. Any discernibility may be achieved only after the fact, by which time the attack has already occurred. In contrast, let’s focus on pre‑emptive measures.

ENISA reports note that attacks by internal adversaries appear in smaller numbers than other threats, though likely with substantial underreporting. For several years, the European Union Agency for Cybersecurity (ENISA) has repeatedly stated in its annual threat reports that

[…] organisations remain reluctant in sharing details of these incidents. [2, 7].

In the 2025 report [1], the issue is scarcely mentioned. Nevertheless, a figure of 0.8% of incidents is reported.

One important note is in order. Given the apparently low figures, one might mistakenly conclude that investments in insider‑threat protections are not economically justified. This is misguided for at least two reasons. First, substantial underreporting likely depresses the figures. One plausible driver is reputational risk: admitting that a rogue insider employee facilitated data exfiltration may be seen as more damaging than attributing the incident to human error—such as an employee clicking a phishing email, which can occur even in well‑trained organizations. Second, as noted earlier, attribution is hard: even after the fact, it is often impossible to determine whether a breach stemmed from a rogue, bribed, or extorted employee, or from human-error that let an external attacker breach the perimeter and possibly impersonate an employee. Many compromises begin with phishing, leaving the internal network only a single click away. It thus remains unclear what fraction of such incidents should be classified as insider attacks. As a result, current statistics likely understate both the prevalence of insider threats and the total damage they cause.

The current solution space for these kinds of threats usually covers many different internal‑surveillance measures that are privacy‑invasive to a company’s employees (see, e.g., [8]). Other solutions revolve around the informal concept of Zero Trust, a security “paradigm” that assumes no user or device—whether inside or outside the network—can be trusted by default, requiring continuous verification and strict access controls for all resources.

However, a provably secure solution that avoids privacy‑invasive measures remains elusive in the broad industry. This is surprising from the standpoint of provable security and cryptographic modeling, where even the simplest threat models typically allow for any party to be corrupted at some point—aside from notable cases that rely on trusted third parties (e.g., a public key infrastructure (PKI) with a root certificate authority (CA)). Protecting against these kinds of threats is hardly a new motivation in the cryptographic literature: work from the 2000s (e.g., [9, 10, 11]) already discusses attacks by internal users, and the idea goes back even further. As early as 1978, Rivest, Adleman, and Dertouzos proposed early approaches to such threats [12]. Today is no different. Consequently, there are nearly five decades of cryptographic research to draw from for such protections.

From my perspective, it appears that when a product is primarily based on a cryptographic protocol, protections of some degree against internal threats within the participating parties are usually in place. The most notable example is the Signal¹ instant‑messaging application, which uses an ingenious double‑ratcheting protocol to quickly recover a client from a potential breach, as analyzed in [13, 14]. Other end‑to‑end‑encrypted file‑transfer tools also mitigate, at a minimum, the attack vector of a compromised back‑end server—for example, croc² or the noisytransfer cli³ from the author of this article. However, adapting cryptographic methods to products that extend beyond simple applications is a non‑trivial task. Currently, multi‑party computation (MPC) appears to be the most versatile tool for achieving such protections, but it suffers from significant overhead due to an overly restrictive adversary model. While an overly restrictive adversary model may seem advantageous, in many cases it goes far beyond what is actually critical within a regular company’s network. This effectively results in unnecessarily large overhead, which hinders MPC solutions from being widely adopted.

From my perspective, modeling an internal adversary should begin with the simplest cases. Protective measures derived from these cases will naturally generalize to more complex scenarios and can often be used as is in other settings. The most basic—though often regarded as uninteresting from a cryptographic standpoint—is an honest employee who unknowingly or mistakenly performs actions that lead to information leakage. This may occur due to misconfiguration or simply not being aware of data‑protection measures in place, and thus unknowingly bypassing them.

Separation of Duties

One may wonder how cryptographic measures can address benign misbehavior. The concept of separation of duties, when incorporated into a cryptographic protocol—such as specific variants of MPC protocols or a contact‑tracing protocol—naturally provides this kind of protection as a byproduct. This concept revolves around distributing distinct processing steps of an application across different systems, operated by separate individuals or (sub‑)organizations. This inherently reduces the risk of complete information disclosure from individual failures by adhering to the principle of not keeping all eggs in one basket.

Passive Confidentiality (i.e., “Honest‑but‑Curious” Security)

In further strengthening defenses against internal adversaries, one may now consider an employee who knowingly breaches security measures with malicious intent—without actively manipulating any systems. Relying solely on the separation‑of‑duties concept will not be sufficient, as such an employee can still leak information—albeit only the information to which they have access to. To ensure confidentiality, encryption methods are required, but this task becomes non‑trivial due to the challenges of key management. Simply encrypting data‑in‑use shifts the internal adversary’s focus from breaching the information itself to breaching the decryption key. This is because, by the nature of data‑in‑use, classically encrypted data must eventually be decrypted to be further processed, and therefore, the decryption key must be accessible and used frequently. This is also why protecting data‑at‑rest against internal adversaries is relatively easier—the decryption key can be kept out of the adversary’s reach, as it is used infrequently. However, Ithe separation‑of‑duties concept can be applied to the key management of data‑in‑use encryption. This would result in one party holding the decryption key and another party processing the encrypted data (i.e., the data‑in‑use). Unfortunately, this approach is not sufficient, as it still requires the encrypted data to be processable. Ideally, data‑in‑use would remain constantly encrypted during processing, with only the final result being decrypted and forwarded to the next processing step, which at that stage might not even require further protection.

This challenge of processing encrypted data was conceptually addressed by the seminal work of [12], which introduced homomorphic encryption (HE). Since then, various encryption schemes have been developed, allowing different degrees of processing on encrypted data. To name a few:

• Fully Homomorphic Encryption (FHE), introduced by [15], allows arbitrary computations on encrypted data but involves significant computational overhead.
• Somewhat Homomorphic Encryption (SHE) [16] permits limited operations on encrypted data and represents a practical trade‑off between functionality and efficiency.
• Deterministic Encryption [17] enables exact search capabilities on encrypted data by generating the same ciphertext for the same plaintext across different encryptions, which is useful for efficient indexing, albeit with reduced semantic security.
• Searchable Encryption [18] facilitates searching encrypted data without revealing the plaintext, a key solution for secure outsourced storage.
• Functional Encryption [19] allows specific computations on encrypted data based on predefined access rights, ensuring selective data access.
• Re‑randomizable Encryption [20] adds an additional layer of security by allowing ciphertexts to be randomized without changing the underlying plaintext.

Digital Operations Resilience Act (DORA [21])

Currently, European regulations in the financial sector explicitly suggest a need for cryptographic measures to protect so‑called data‑in‑use. See, for example, Article 6 Encryption and Cryptographic Controls:

2. The policy on encryption and cryptographic controls shall be designed based on the results of approved data classification and ICT risk assessments, and shall include the following elements: […] (b) rules for the encryption of data-in-use, where necessary. […] [21]

As already mentioned, traditional encryption methods—such as those that satisfy strong security standards like chosen‑ciphertext attack (CCA)—are not suitable for encrypting data‑in‑use, as they render the data unusable for subsequent processing. Thus, it appears that these regulations may be encouraging the use of advanced techniques like homomorphic encryption (HE) or multi‑party computation (MPC) for encrypting data‑in‑use, thereby strengthening financial systems against insider threats.

Pseudonymization

Alternative means of protecting personal data‑in‑use are pseudonymizations. The current regulation and best‑practices landscape provides plenty of resources (cf. [22, 23, 24, 25, 26]). I have provided an extensive summary on what Pseudonymization means. Roughly, Pseudonymization is a data‑protection technique that replaces identifiable information within a dataset with pseudonyms and de‑pseudonymizes them when necessary for processing. This process reduces the risk of re‑identification while allowing the data to remain useful for analysis and processing. According to the General Data Protection Regulation (GDPR) Article 4(5), pseudonymization is defined as:

The processing of personal data in such a manner that the personal data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately and is subject to technical and organizational measures to ensure that the personal data are not attributed to an identified or identifiable natural person. [27]

There are also other definitions available from ISO/IEC 20889 [28] and ISO 25237 [29] standards. However, all definitions share the essence of protecting personal data while keeping it usable, which effectively means that they are data‑in‑use protections against internal adversaries as well. To emphasize this observation, consider paragraph 42 of the guidelines on pseudonymization from the European Data Protection Board (EDPB):

If a controller or processor wants to use pseudonymisation to reduce confidentiality risks from some or all unauthorised third parties, they will include those third parties in the pseudonymisation domain and assess the means they are reasonably likely to use for attribution. Relevant third parties include not only cyber-crime actors, but also employees or maintenance service providers acting in their own interests rather than on instructions from the controller. Taking due account of contextual elements and the circumstances at hand, it is recommended to consider both actions in good faith, and those executed with criminal intent. [25]

Although the regulatory definitions are valuable, they are not formal in the sense of provable security; they require thorough risk analysis and familiarity with state‑of‑the‑art pseudonymization techniques for effective implementation, which can be categorized in the following way:

• Pseudonymization: Involves replacing identifiable data with pseudonyms using symmetric cryptographic techniques.
• Delegated Pseudonymization: Involves outsourcing or delegating the pseudonymization process to someone else—typically the data subjects themselves or a third party holding the data. These methods involve the use of asymmetric cryptographic techniques. Usually, these techniques are part of the privacy‑enhancing technology (PET) or privacy‑enhancing cryptography (PEC) landscape [30].

Thus, the previously mentioned encryption schemes—such as FHE, SHE, deterministic encryption, functional encryption, and so on—are considered pseudonymization techniques in the industrial state of the art. Therefore, I arrive at the same cryptographic foundation when considering pseudonymization as is the case with the data‑in‑use protections in DORA [21].

A Note on Privacy and Security

At the outset, I argued that protections against internal attackers are security measures, whereas pseudonymization methods are usually portrayed as privacy measures. I contend, however, that pseudonymization techniques are precisely the same mechanisms used to protect data‑in‑use against internal threats. This overlap exists because measures that safeguard individuals’ personal data also serve as security controls for the organization holding that data, defending it from both external and internal adversaries. Consequently, if data processing is not secure against internal attackers, privacy cannot be achieved. In other words, privacy for data subjects is security for the data controller.

Outlook on Privacy & Security in the Real World

Despite significant advances in regulatory frameworks and best‑practice recommendations, privacy‑enhancing methods have yet to achieve broad industrial adoption. Waldman [31] offers several plausible explanations for why privacy—and, consequently, the techniques designed to protect it—remain niche concerns (if acknowledged at all) in industry. As a layperson in the field of privacy law, I am not positioned to offer a definitive assessment of these legal arguments, nor to propose specific amendments to the General Data Protection Regulation to address this lack of methodological uptake.

Last but not least there is also the topic of Confidential Computing, which adresses the same prolem. But i will leave this discussion for the future.

References

1. ENISA. (2025). Cyber Threats 2025. https://www.enisa.europa.eu/publications/enisa-threat-landscape-2025
2. ENISA. (2024). Cyber Threats 2024. https://www.enisa.europa.eu/topics/cyber-threats/threats-and-trends
3. BSI. (2024). The State of IT Security in Germany. https://www.bsi.bund.de/EN/Service-Navi/Publikationen/Lagebericht/lagebericht_node.html
4. Deloitte. (2024). Cybersecurity Threat Trends Report 2024. https://www2.deloitte.com/us/en/pages/risk/articles/cybersecurity-threat-trends-report-2024.html
5. Thales. (2024). Data Threat Report. https://cpl.thalesgroup.com/data-threat-report
6. Truesec. (2024). Truesec Threat Intelligence Report 2024. https://insights.truesec.com/hub/report/truesec-threat-intelligence-report-2024
7. ENISA. (2024). Cyber Threat Landscape 2024. https://www.enisa.europa.eu/publications/enisa-threat-landscape-2024
8. CISA. (2020). Insider Threat Mitigation Guide. https://www.cisa.gov/resources-tools/resources/insider-threat-mitigation-guide
9. Ge, T. & Zdonik, S. (2007). Answering aggregation queries in a secure system model. In Proceedings of the 33rd international conference on Very large data bases, pp. 519—530.
10. Halevi, S., Lindell, Y., & Pinkas, B. (2011). Secure computation on the web: Computing without simultaneous interaction. In Advances in Cryptology—CRYPTO 2011: 31st Annual Cryptology Conference, Santa Barbara, CA, USA, August 14-18, 2011. Proceedings 31, pp. 132—150.
11. Gordon, S. D., Malkin, T., Rosulek, M., & Wee, H. (2013). Multi-party computation of polynomials and branching programs without simultaneous interaction. In Advances in Cryptology—EUROCRYPT 2013: 32nd Annual International Conference on the Theory and Applications of Cryptographic Techniques, Athens, Greece, May 26-30, 2013. Proceedings 32, pp. 575—591.
12. Rivest, R. L., Adleman, L., & Dertouzos, M. L. (1978). On Data Banks and Privacy Homomorphisms. In Foundations of secure computation, pp. 171—189, “Academic Press.
13. Bienstock, A., Fairoze, J., Garg, S., Mukherjee, P., & Raghuraman, S. (2022). A more complete analysis of the signal double ratchet algorithm. In Annual International Cryptology Conference, pp. 784—813.
14. Cremers, C., Medinger, N., & Naska, A. (2025). Impossibility Results for Post-Compromise Security in Real-World Communication Systems. In 2025 IEEE Symposium on Security and Privacy (SP), pp. 4391—4405.
15. Gentry, C. (2009). Fully homomorphic encryption using ideal lattices. In Proceedings of the forty-first annual ACM symposium on Theory of computing, pp. 169—178.
16. Van Dijk, M., Gentry, C., Halevi, S., & Vaikuntanathan, V. (2010). Fully homomorphic encryption over the integers. In Advances in Cryptology—EUROCRYPT 2010: 29th Annual International Conference on the Theory and Applications of Cryptographic Techniques, French Riviera, May 30—June 3, 2010. Proceedings 29, pp. 24—43.
17. Bellare, M., Boldyreva, A., & O’Neill, A. (2007). Deterministic and efficiently searchable encryption. In Advances in Cryptology-CRYPTO 2007: 27th Annual International Cryptology Conference, Santa Barbara, CA, USA, August 19-23, 2007. Proceedings 27, pp. 535—552.
18. Song, D. X., Wagner, D., & Perrig, A. (2000). Practical techniques for searches on encrypted data. In Proceeding 2000 IEEE symposium on security and privacy. S&P 2000, pp. 44—55.
19. Boneh, D., Sahai, A., & Waters, B. (2011). Functional encryption: Definitions and challenges. In Theory of Cryptography: 8th Theory of Cryptography Conference, TCC 2011, Providence, RI, USA, March 28-30, 2011. Proceedings 8, pp. 253—273.
20. Canetti, R., Halevi, S., & Katz, J. (2003). A forward-secure public-key encryption scheme. In Advances in Cryptology—EUROCRYPT 2003: International Conference on the Theory and Applications of Cryptographic Techniques, Warsaw, Poland, May 4—8, 2003 Proceedings 22, pp. 255—271.
21. ESA. (2024). Final report on draft RTS on ICT Risk Management Framework and on simplified ICT Risk Management Framework. https://www.eba.europa.eu/sites/default/files/2024-01/bf5a2976-1a48-44f3-b5a7-56acd23ba55c/JC%202023%2086%20-%20Final%20report%20on%20draft%20RTS%20on%20ICT%20Risk%20Management%20Framework%20and%20on%20simplified%20ICT%20Risk%20Management%20Framework.pdf
22. ENISA. (2019). Pseudonymisation Techniques and Best Practices. https://www.enisa.europa.eu/publications/pseudonymisation-techniques-and-best-practices
23. ENISA. (2021). Data Pseudonymisation: Advanced Techniques and Use Cases. https://www.enisa.europa.eu/publications/data-pseudonymisation-advanced-techniques-and-use-cases
24. ENISA. (2022). Deploying Pseudonymisation Techniques. https://www.enisa.europa.eu/publications/deploying-pseudonymisation-techniques
25. EDPB. (2025). Guidelines 01/2025 on Pseudonymization. https://www.edpb.europa.eu/our-work-tools/documents/public-consultations/2025/guidelines-012025-pseudonymisation_en
26. EP. (2018). Recommendations on shaping technology according to GDPR provisions. https://op.europa.eu/en/publication-detail/-/publication/0e1ca64f-29c7-11e9-8d04-01aa75ed71a1/language-en
27. EP. (2016). Art. 5 GDPR Principles relating to processing of personal data. https://gdpr-info.eu/art-5-gdpr/
28. ISO/IEC. (2018). ISO/IEC 20889:2018 Privacy enhancing data de-identification terminology and classification of techniques. https://www.iso.org/standard/69373.html
29. ISO. (2017). ISO 25237:2017 Health informatics — Pseudonymization. https://www.iso.org/standard/63553.html
30. NIST. (2024). NIST Privacy-Enhancing Cryptography Project. https://csrc.nist.gov/projects/pec
31. Waldman, A. E. (2019). Privacy law’s false promise. Wash. UL Rev., 97, 773.