Confidential Computing Explained: How TEEs Protect Data in Use
Confidential computing is a security approach that protects data in use with hardware-level TEEs. Through isolated environments, memory encryption, and remote attestation, it reduces the risk of unauthorized access and is suitable for sensitive data AI and multi-party data analysis.
What Is Confidential Computing?
Confidential Computing is a method for protecting data in use through hardware-based trusted execution environments, or TEEs. It focuses on the security state of data while it is being processed by applications, analytics tasks, model inference, or machine learning workloads—not only on the security of data when it is stored or transmitted.
In traditional security systems, data is usually protected through encrypted channels during transmission and through encryption at rest during storage. But when data enters memory and is processed by a program, system administrators, host software, or other highly privileged components may still become potential risk points. The gap that confidential computing aims to fill is precisely this protection gap for “data in use.”
Key point: Confidential computing does not simply lock data away. Instead, it allows data to be computed inside a protected hardware-isolated environment, thereby reducing the risk of unauthorized access or modification.
The Core Problem Confidential Computing Solves
The core problem confidential computing aims to solve can be summarized as follows: when data must be computed, how can its exposure to the runtime environment, operations personnel, and other software components be minimized as much as possible?
“Data in use” commonly appears in the following scenarios:
- An application service is processing user requests or business data;
- An analytics task is computing on sensitive datasets;
- An AI model is performing inference or training;
- Multi-party data collaboration requires joint analysis, but each party does not want to expose its raw data.
What these scenarios have in common is that data cannot remain statically encrypted forever; it must enter the runtime environment and participate in computation. Confidential computing creates a protected execution area through a hardware-level TEE, allowing applications and data to run in an isolated environment and reducing the possibility that external systems can directly read, tamper with, or inspect runtime data.
How a Trusted Execution Environment TEE Provides Isolation Protection
A TEE, or trusted execution environment, is a secure and independent execution environment. Its role is to prevent unauthorized access to or modification of applications and data in use.
In confidential computing, a TEE usually undertakes three key responsibilities:
| Mechanism | Role | Problem Solved |
|---|---|---|
| Hardware isolation | Isolates protected workloads from the ordinary host environment | Reduces the risk that host software or other components directly come into contact with sensitive data |
| Controlled execution | Allows applications to process data inside a secure area | Prevents runtime data from being read or modified without authorization |
| Reduced trust boundary | Limits the trusted scope to the hardware-supported secure environment | Reduces dependence on the external operations environment |
A TEE can be understood as a “safe in which files can be processed.” An ordinary safe is only responsible for storing files, while the goal of confidential computing is to allow files to be read, computed, and processed inside the safe. Even if external personnel manage the room, they should not directly see the contents inside the safe.
The isolation goal of confidential computing is not only ordinary attackers, but also system administrators or software components that may have higher privileges. In other words, it attempts to reduce this risk: the fact that the host environment has administrative privileges does not necessarily mean it can see the data inside the secure area.
A Typical Confidential Computing Workflow
The implementation of confidential computing varies by platform and hardware, but a typical workflow usually includes several stages: establishing a protected environment, runtime isolation, memory protection, and trustworthiness verification.
Workloads Enter the Protected Environment
Applications, models, or data processing tasks are placed into a hardware-based TEE to run. This environment is isolated from the host system operations environment and is used to carry workloads that need protection and data in use.
At this stage, the core objective is to first establish a clear security boundary: which code and data run inside the protected area, and which components are outside the boundary.
Runtime Data Is Processed in the Isolated Environment
After the workload starts executing, sensitive data is processed inside the isolated environment. The TEE restricts external systems, host software, or other unauthorized components from accessing or modifying applications and data inside the secure area.
This step corresponds to the core value of confidential computing: protecting the process in which data is being used, rather than only protecting the state of the data before it enters the system or after processing is completed.
Hardware Root of Trust and Boot Chain Measurement Confirm That the Environment Is Trusted
Some confidential computing implementations are based on a hardware root of trust and use boot chain measurement to confirm that the runtime environment has been in the expected state since the boot stage. The significance of boot chain measurement is that it not only protects data in operation, but also tries to confirm that the environment carrying this data has not deviated from expectations during the boot process.
Such mechanisms provide a foundation for subsequent trusted execution and verification.
Memory Encryption Protects the Confidentiality of Runtime Data
Memory encryption is used to protect the confidentiality of runtime data. Since data in use usually appears in memory, memory-side protection can reduce the risk of data exposure during the runtime stage.
It should be noted that memory encryption is not equivalent to a complete application security system. It is one of the important mechanisms in confidential computing and usually needs to be used together with capabilities such as TEE isolation, boot chain measurement, and access control.
Remote Attestation Provides Verifiable Trust
Remote attestation is used to allow external participants to verify whether a workload is running in the expected trusted environment. For multi-party data analysis, cross-organization collaboration, or sensitive data AI scenarios, remote attestation is important because participants often cannot rely only on the other party’s verbal commitment.
Through remote attestation, external systems can confirm that the runtime environment has the expected trusted attributes before exchanging sensitive data or initiating computing tasks.
Confidential Computing Differs from Common Data Security Measures in Focus
Confidential computing is not a replacement for encryption in transit or storage encryption. Instead, it supplements protection for the stage of “data in use.” In actual systems, multiple security measures usually need to be combined to cover different stages of the data lifecycle.
| Security Measure | Main Protection Stage | Main Goal | Typical Mechanisms | Applicable Scenarios |
|---|---|---|---|---|
| Encryption in transit | During data transmission | Reduces the risk of eavesdropping and tampering during network transmission | Encrypted channels, certificate validation, etc. | API calls, service-to-service communication, client-server communication |
| Storage encryption | When data is stored at rest | Reduces the risk of data exposure after disks, object storage, or backup media are leaked | Disk encryption, file encryption, key management, etc. | Databases, file storage, backup archiving |
| Confidential computing | While data is in use | Reduces the risk that runtime data is accessed or modified without authorization | Hardware-level TEEs, isolated environments, memory encryption, remote attestation | Sensitive data computation, multi-party analysis, AI workloads |
If only encryption in transit and storage encryption are used, data may still face runtime exposure risks after it enters the application runtime environment. The value of confidential computing lies in pushing the security boundary to where computation takes place, so that sensitive data remains protected while it is being processed.
Application Value in Multi-Party Data Analysis and Machine Learning
Confidential computing is especially suitable for scenarios in which data must be used jointly, but participants do not want to directly expose raw data. Multi-party data analysis, machine learning, and sensitive data AI are typical application directions.
| User Need | Solution | Effect |
|---|---|---|
| Multiple organizations want to jointly analyze data but do not want to expose raw data to one another | Place analytics or machine learning workloads into a TEE-based isolated environment to run | Combined datasets can be computed while maintaining privacy among participants |
| An enterprise wants to run AI on sensitive data | Use confidential computing to protect the model and data in use | Reduces the risk of sensitive data exposure during model inference or training |
| Participants need to confirm that the computing environment is trusted | Use remote attestation to verify that the workload runs in the expected trusted environment | Provides a verifiable foundation of trust for cross-organization collaboration |
| The business needs to reduce the risk of highly privileged access | Restrict privileged software or personnel from accessing data inside the secure area by isolating the workload environment | Reduces direct reliance on trust in the host operations environment |
In AI scenarios, confidential computing can be used not only to protect input data, but also to protect models and the computation process during runtime. For privacy-sensitive analytics, cross-organization data collaboration, model inference, or training tasks, it allows data to participate in computation while maintaining controlled exposure.
Applicable Conditions to Confirm Before Introducing Confidential Computing
Confidential computing is more suitable for businesses that truly need runtime data protection. Before introducing it, you can first determine whether it is a match from three aspects: security objectives, collaboration model, and platform capabilities.
Applicability Checklist
- Is there a situation where sensitive data must enter an application, analytics task, or AI workload to be processed?
- Do you want to reduce the risk that system administrators, highly privileged software, or the host environment can access runtime data?
- Is there a need for multi-party data collaboration where each party does not want to directly share raw data?
- Is there a need to prove to external participants that the workload is running in the expected trusted environment?
- Does the selected platform or solution support hardware-level TEEs?
- Are key mechanisms such as memory encryption, boot chain measurement, and remote attestation available?
- Has it already been made clear which data, code, and models need to enter the protected environment?
Scenarios More Suitable for Priority Evaluation
If the business focus is sensitive data computation, multi-party collaboration, or AI workloads, confidential computing is usually more worth evaluating as a priority. For example, scenarios such as cross-organization joint analysis, privacy-sensitive model inference, and sensitive data processing may all require additional protection during the data usage stage.
Factors That Need Careful Verification
The actual implementation effect of confidential computing is affected by hardware, cloud platforms, operating systems, runtimes, application architecture, and workload types. Performance overhead, cost, compatibility, and deployment complexity should not be concluded in advance without actual testing.
Key point: Confidential computing solves the question of “how data in use is protected.” It should be used as part of an overall security architecture, together with identity authentication, access control, encryption in transit, storage encryption, key management, and audit mechanisms.