Abstract
How can one tamper with data that does not exist? Motivated by this question, we present the Burning Fetch eXecution (BFX) paradigm. Data in-use is vulnerable, and the current focus on encrypting and/or isolating in-use data has fallen short. Frequently reported breaches of “secure” hardware and indispensable overhead with encryption schemes confirm that trust is the modern bottleneck. This work tackles the gap in existing safeguarding technology by avoiding byte-level decryption until it is immediately fetched by the processor, only to burn it right after. We perform on-the-fetch data decryption, immediately followed by burning, i.e., erasing right after processing cycles. Thus, BFX minimizes the existence of sensitive data in-use. BFX does not demand new processing hardware units nor requires restructuring application software. Three pillars set the BFX paradigm apart: (1) zero-trust multi-party confidentiality with (2) security rooted in transparency, and (3) high performance.
The Vulnerability of Data In-Use
In modern cybersecurity, protecting data at-rest (on a hard drive) and in-transit (moving over a network) is a solved problem through robust encryption. However, data in-use—the information actively being processed by a CPU—remains a critical vulnerability.
Traditional systems must decrypt data into memory to process it, creating a window of opportunity for "insider threats" (like a rogue cloud admin) or side-channel attacks. While Trusted Execution Environments (TEEs) like Intel SGX aim to address this, they often suffer from hardware vulnerabilities and significant performance overhead. BFX addresses this by ensuring that plaintext data exists for only a few clock cycles, effectively shrinking the attack surface to near zero.
The Burning Fetch eXecution (BFX) Paradigm
The BFX framework rests on three technical pillars designed to eliminate the need for trust in the underlying hardware or cloud provider:
- Just-on-Fetch: Data remains encrypted in main memory at all times. It is decrypted at the hardware level only when the processor specifically requests a memory line, ensuring only a tiny sliver of data is ever exposed.
- Fetch-n-Burn: Immediately after the processor completes its cycles with the data, the BFX hardware instantly and irreversibly erases the decrypted bytes.
- Fetch-n-Monitor: The system tracks fetches in real-time, detecting and shutting down unauthorized access attempts by compromised operating systems or rogue processes.
By interposing a programmable module (such as an FPGA) between the processor and memory, BFX works with standard off-the-shelf (COTS) hardware without requiring software rewrites or chip redesigns.
Multi-Party Zero-Trust
BFX enables a "sharing without showing" model. In a typical scenario, a Data Owner (e.g., a drone operator) and a Code Owner (e.g., an AI developer) may need to use a third-party Data Processor (e.g., a cloud provider). BFX ensures that even if the cloud provider is compromised, the attacker cannot access the sensitive data or the proprietary code because the hardware enforcement layer prevents plaintext from persisting in memory.
Evaluation: The Drone Swarm Mission
To demonstrate BFX, the authors simulated a real-world mission involving a live, encrypted video stream from a drone swarm. A "top-secret" TensorFlow-Lite neural network was used for real-time object detection within a compromised cloud environment where an attacker had root access.
Key Results
- Security against Privileged Attackers: While a "sudoer" attacker could easily dump memory and view the video feed in a standard setup, BFX rendered the memory as randomized, encrypted bytes.
- Real-Time Performance: The system maintained real-time inference speeds on the live feed, proving that high security does not necessitate the massive performance penalties (sometimes 400x) seen in other encryption schemes.
- Transparency: Unlike proprietary "black box" TEEs, BFX's programmable layer allows for auditability and verification by all parties involved.
Conclusion
BFX shifts the security paradigm from isolation to elimination. By ensuring that sensitive data effectively "does not exist" except during the exact cycles of computation, it provides a high-performance, zero-trust solution for the most demanding multi-party computing environments. Future work will focus on extending this protection to specialized accelerators like GPUs and scaling the architecture for data-center-class memory bandwidth.
Summary updated