Introduction
As sensitive data increasingly migrates to cloud environments, two critical security requirements emerge: data confidentiality and computational integrity. Confidentiality ensures that the cloud provider cannot access raw user data, while integrity guarantees that the provider correctly executes requested computations.
Current solutions typically prioritize one over the other. Homomorphic Encryption (HE) allows for mathematical operations on encrypted data, preserving privacy but suffering from "malleability"—where an attacker can modify ciphertexts to predictably alter underlying values without detection. Conversely, Trusted Execution Environments (TEEs), such as Intel SGX, provide hardware-isolated enclaves for secure computation. While TEEs offer strong integrity, they require users to place absolute trust in the hardware manufacturer; a hardware compromise leads to a total loss of privacy.
This paper proposes a "combined model" that pairs Partially Homomorphic Encryption (PHE) with TEEs to decouple data confidentiality from computational integrity, effectively mitigating the weaknesses of both individual approaches.
Methodology: The Combined Model
The proposed architecture utilizes a TEE to safeguard the inherent malleability of HE. The system is built on two primary layers:
- Encryption Layer: Data is protected using the Paillier cryptosystem, a form of PHE that supports additive operations. This ensures that even if the TEE hardware is compromised, the underlying data remains encrypted and unreadable to the provider.
- Integrity Layer: Computation logic resides within an Intel SGX enclave. This enclave manages database keys and verifies authentication tags (CMAC) on the data to prevent unauthorized modifications.
In this protocol, the user encrypts data with a PHE key and wraps it in a network-encrypted layer before transmission. The enclave strips the network layer but does not decrypt the PHE layer. It then performs homomorphic operations (e.g., summing encrypted values) and attaches an integrity check. The final result is returned to the user, who holds the sole key for decryption. This approach ensures that the TEE never sees the raw data, only the encrypted ciphertexts it is operating upon.
Performance Evaluation
The authors evaluated the model using a simulated database containing 1,000 entries. Each entry consisted of a 64-bit integer stored within a 2048-bit container to accommodate Paillier ciphertexts. Testing was conducted on a 7th Gen Intel Core desktop environment.
The results indicate that the combined model is approximately 1.7x slower than using either SGX or PHE in isolation. However, the execution time scales linearly with the number of database entries, suggesting that the overhead remains manageable for practical applications. For a summation query, the performance penalty is constant relative to the operation's complexity.
Key Insights and Discussion
The most significant contribution of this work is the decoupling of trust. By using math (HE) for privacy and hardware (TEE) for integrity, users no longer need to trust hardware vendors with raw data. The TEE's role is reduced to ensuring mathematical correctness and blocking malleability attacks by verifying authentication tags before processing.
While the current implementation is limited to additive operations due to the use of Paillier PHE, it remains highly practical for use cases like electronic voting or financial auditing. In a voting scenario, the model allows a TEE to aggregate votes homomorphically, maintaining voter anonymity without requiring complex Multi-Party Computation (MPC) protocols.
Limitations and Future Work
The primary limitation is functional; complex logic like searching or multiplication would require moving to Fully Homomorphic Encryption (FHE), which carries a much higher performance cost. Additionally, while the model protects privacy during a TEE breach, the user still relies on the hardware for the correctness of the result. Future research will explore hardware optimizations for large-integer math and the application of this model to more complex protocols such as secure auctions.
Summary updated