A Paradigm Shift in Biometric Information Protection with Fully Homomorphic Encryption
Picture this: You walk through airport security, and your face alone unlocks the doors. Behind the scenes, systems silently verify your identity against official documents—no fumbling for IDs, no lines, no hassle.
A recent CNN article, The Race to Become the World’s First Document-Free Airport [1], paints this exact future. By 2025, Abu Dhabi’s Zayed International Airport plans to make this vision a reality. Similar innovations are already in motion at Denver International Airport in the U.S. [2] and Incheon International Airport in South Korea [3].
Facial recognition is not only transforming how we travel—it’s also reshaping how we pay. According to CNBC, biometric payment methods are rapidly gaining ground in the U.S., with Apple Face ID enabling seamless app purchases. Financial leaders like JPMorgan and Mastercard are following suit [4]. Their report, With JPMorgan, Mastercard on Board in Biometric ‘Breakthrough’ Year, You May Soon Start Paying with Your Face, underscores the accelerating adoption of facial recognition in finance.
Beyond airports and payments, facial recognition is already embedded in our daily lives—from building access and entertainment venue entry to law enforcement and border control. As the technology matures, its adoption will likely continue to expand across industries.
But with this rapid growth comes a growing set of privacy and security concerns. Unlike passwords, biometric data such as facial templates cannot be changed if stolen. A breach could have serious, long-lasting consequences.
That’s where Fully Homomorphic Encryption (FHE) comes in—a cutting-edge cryptographic technique that allows computations to be performed directly on encrypted data. CryptoLab’s Encrypted Facial Recognition (EFR) solution leverages FHE to ensure that facial templates remain encrypted at all times—even during identity verification or matching. Sensitive biometric information never needs to be decrypted, greatly reducing the risk of exposure.
This blog dives into how FHE works and explores the key advantages of CryptoLab’s EFR over traditional data protection methods—offering a powerful, privacy-preserving alternative for the age of biometric recognition.
How facial recognition works
Facial recognition operates through advanced algorithms and machine learning to identify and verify individuals based on distinct facial landmarks such as the eyes, nose, mouth, and jawline. When an image is captured, the facial recognition system generates a face template—a numerical representation that encodes unique facial features. This template is typically stored as a high-dimensional vector, often in the form of single-precision floating-point values (e.g., 512-length vectors), allowing precise differentiation between individuals.
The system then compares this template against a database to determine a match, typically using cosine similarity or Euclidean distance. In a One-to-One match, used for authentication (e.g., unlocking a phone), the system verifies an input facial template against a stored template. In contrast, One-to-Many identification, common in airports and pay-by-face systems, involves comparing an input facial template against a database of templates to find a match.
Facial Data Reconstruction
At first glance, reconstructing a facial image from a template may seem impossible — after all, these templates are merely numerical vectors. But thanks to advances in machine learning and the rich facial landmark data they encode — enough to distinguish one face among millions — reversing facial templates is becoming increasingly feasible. A recent study demonstrates that images reconstructed from facial templates closely resemble the original, as shown below [8].
Therefore, safeguarding facial templates is essential. This requires protection at all stages — at-rest, in-transit, and in-use. However, conventional cryptographic methods do not provide in-use protection, since facial templates must be decrypted for matching. This is where fully homomorphic encryption offers a breakthrough.
Fully Homomorphic Encryption
Homomorphic encryption is a cryptographic technique that allows computations to be performed on encrypted data without decrypting it first. The idea was originally proposed in 1978 by Ron Rivest, Leonard Adleman, and Michael Dertouzos in their paper On Data Banks and Privacy Homomorphisms. Rivest and Adleman are two of the three inventors of RSA encryption.
For the following 30 years, cryptographers around the world have embarked on a quest to construct an encryption scheme that would enable arbitrary computation on encrypted data. They were able to come up with homomorphic encryption schemes that supported a limited number or types of operations, but a scheme that allows unlimited arbitrary operations on encrypted data, coined “Fully Homomorphic Encryption”, remained elusive until Craig Gentry, a graduate student at Stanford University, introduced the first construction of a FHE scheme in his doctoral thesis in 2009.
Four generations of FHE with their applications
FHE is revolutionizing privacy and security by enabling computations on encrypted data without decryption while allowing unlimited arbitrary computations to be carried out. This means any mathematical operation that can be applied to plaintext can also be performed on encrypted data, preserving both functionality and privacy.
But how can computations be performed on encrypted data? Although homomorphic encryption schemes used in the real world—including CryptoLab’s solution—are based on complex mathematical theories, we will explore the concept here using a greatly simplified example for better understanding. In this simplified homomorphic encryption scheme, numbers are transformed into ciphertexts through an encryption process and then returned to their original values through a decryption process.
Encryption: plaintext + (key x random)
Decryption: ciphertext mod key
Encrypted Facial Match
To safeguard facial templates, it is essential to secure both the database and the face matching process, ensuring that plaintext data never appears in memory. Homomorphic encryption can protect the facial matching process, as it relies on operations such as addition and multiplication—which homomorphic encryption supports. For instance, cosine similarity, a common method for face matching, is computed as follows:
Since homomorphic encryption supports these operations, it can effectively be used to calculate the cosine similarity of two encrypted ciphertexts without exposing plaintext data.
Encrypted Facial Recognition (EFR) by CryptoLab
Encrypted Facial Recognition (EFR) is CryptoLab’s solution for protecting facial templates using 4th-generation fully homomorphic encryption scheme, CKKS. In EFR, facial templates are encrypted at all times – they are stored and processed encrypted. During one-to-many identification, an input template is compared against stored encrypted templates without decryption.
Key strengths of EFR include:
It is immune to data breaches, side-channel attacks, supply chain risks, insider threats, and a range of other cyber and privacy threats, because the data is always encrypted even while being computed upon.
Because the data is always encrypted, EFR database and EFR computations can be securely hosted in off-prem hosts.
EFR performs encrypted facial recognition in a fraction of a second against tens of millions of encrypted face template vectors with GPU acceleration. Additionally, EFR maintains performance with additional GPUs – estimated 1 to 2 GPUs per million face templates.
EFR accuracy is comparable to plaintext facial recognition results.
EFR provides quantum resistant data protection – CKKS is based on Lattice cryptography like NIST’s ML-KEM and ML-DSA.
CryptoLab’s EFR is available for commercial licensing to enterprises and developers. For more details about EFR, please visit https://fhefr.com/.
Reference
[1] CNN, DeOliva, Ana. “The race to become the world’s first document-free airport.”, 7 August 2024, https://edition.cnn.com/travel/abu-dhabi-smart-travel-project/index.html.
[2] Denver International Airport, “Denver International Airport Installs Biometric Boarding Devices for International Departures.” https://www.flydenver.com/press-release/denver-international-airport-installs-biometric-boarding-devices-for-international-departures/.
[3] JoongAng Daily, “Forget the passport, your face is all you need in Korea.” Korea, 6 September 2023, https://koreajoongangdaily.joins.com/news/2023-09-06/business/industry/Forget-the-passport-your-face-is-all-you-need-in-Korea/1863546.
[4] CNBC, “The world’s first airport to require biometric boarding is set to arrive in 2025.” 22 August 2024, https://www.cnbc.com/2024/08/22/worlds-first-airport-to-require-biometric-boarding-to-arrive-in-2025.html
[5] MIT News, Michalowski, Jennifer. “An optimized solution for face recognition.” MIT News, 6 April 2022, https://news.mit.edu/2022/optimized-solution-face-recognition-0406. Accessed 22 October 2024
[6] NIST, “Face Technology Evaluations – FRTE/FATE, “ https://www.nist.gov/programs-projects/face-technology-evaluations-frtefate
Report: https://pages.nist.gov/frvt/reports/1N/frvt_1N_report.pdf
[7] H. Otroshi Shahreza, V. K. Hahn and S. Marcel, “Vulnerability of State-of-the-Art Face Recognition Models to Template Inversion Attack,” in IEEE Transactions on Information Forensics and Security, vol. 19, pp. 4585-4600, 2024
