Building the Truman Show: A TrustZone-Based Framework for Lightweight Out-of-band Kernel Security Monitoring

Zhenling Duan; Pan Dong; Renshuang Jiang; Xiaoxiang Fang; Bao Li

Source-linked AI summary

Building the Truman Show: A TrustZone-Based Framework for Lightweight Out-of-band Kernel Security Monitoring

Zhenling Duan, Pan Dong, Renshuang Jiang, Xiaoxiang Fang, Bao Li

arXiv:2608.29758v1Aug 30, 2026cs.CRcs.SE

TL;DR

Kernel-resident defenses can fail with the kernel, while strict isolation limits semantic visibility and timely monitoring. LOOM uses ARM TrustZone, selective Secure World semantic reconstruction, and layered prevention to monitor kernel threats out of band. Its prototype reports negligible performance overhead, while its scope remains focused on specific rootkit-related kernel objects and excludes several attack classes.

Problem
Kernel vulnerabilities and rootkits undermine in-kernel defenses, while isolation-based protection can create semantic gaps that limit kernel visibility and timely detection.
Method
LOOM uses TrustZone-isolated Secure World monitoring with selective reconstruction of critical kernel-object attributes and layered hazard prevention.
Results
LOOM’s prototype and evaluation report negligible performance overhead while demonstrating detection capabilities through representative CVE cases.
Takeaways & Limitations
Out-of-band TrustZone monitoring provides a kernel-independent architecture for analyzing and responding to selected rootkit-related kernel behaviors.
Takeaways & Limitations
LOOM targets security-critical kernel objects associated with specific rootkit surfaces and excludes physical hardware, side-channel, and denial-of-service attacks.

Abstract

from arXiv · show

The increasing number of vulnerabilities in operating systems, together with sophisticated kernel-level threats (e.g., rootkits), has weakened the effectiveness of traditional in-kernel protection mechanisms. Since these defenses operate at the same privilege level as the kernel, they share the same attack surface and can be bypassed once the kernel is compromised. Isolation-based security approaches provide stronger protection by separating security logic from the kernel, but strict isolation often introduces semantic gaps that limit system visibility and hinder timely threat detection. In this paper, we present LOOM, a lightweight out-of-band operating system monitoring architecture built on ARM TrustZone. By leveraging TrustZone's hardware-enforced isolation, LOOM establishes a tamper-resistant monitoring environment independent of the kernel. To bridge the semantic gap, we design a lightweight semantic reconstruction mechanism in the Secure World. It selectively captures the states and behavioral patterns of critical kernel objects, such as process control blocks and kernel modules. Additionally, LOOM introduces a dual-stage hazard prevention mechanism that combines atomic memory protection with an interrupt-driven adaptive agent to detect and mitigate kernel rootkit activities. An address translation cache is further incorporated to optimize repeated address access and reduce monitoring overhead. Overall, we develop a multi-layered collaborative architecture with platform, functional, and auxiliary layers for secure and efficient kernel monitoring. A prototype of LOOM has been implemented on the Phytium D2000 platform. Experimental results indicate that LOOM incurs negligible overhead while maintaining a strong monitoring capability. Furthermore, a security capability analysis based on CVE cases demonstrates that LOOM can detect and mitigate various kernel attacks.

1 Introduction

LOOM addresses the weakness of kernel-resident protection by moving monitoring into an isolated TrustZone Secure World while selectively reconstructing kernel semantics. Its layered design combines out-of-band monitoring, hazard prevention, and performance optimization for kernel security.

Motivation: Kernel-resident defenses share the kernel’s attack surface and can fail after kernel compromise.Kernel complexity also exposes in-kernel mechanisms to privilege escalation and control-flow hijacking.
Motivation: Isolation improves security independence but creates semantic gaps that reduce system visibility and complicate timely detection.The paper identifies challenges in understanding kernel semantics, balancing reconstruction completeness against complexity, and responding quickly to threats.
Approach: LOOM uses ARM TrustZone to place tamper-resistant monitoring in a Secure World separate from the potentially compromised Normal World kernel.TrustZone also supports low-intrusion memory operations through one-way access and privileged register interfaces.
Approach: A Secure World semantic reconstruction module selectively captures states and behavior patterns of critical kernel objects, including process control blocks, file descriptor tables, and module lists.The design avoids reconstructing the entire kernel while supporting security analysis of targeted objects.
Contributions: LOOM combines a three-layer architecture with dual-stage hazard prevention and an address translation cache to support secure, efficient monitoring.The platform, functional, and auxiliary layers respectively provide isolation and access, semantics-aware monitoring and control, and overhead optimization.
Evaluation: A Phytium D2000 prototype and CVE case studies show negligible overhead and detection of representative kernel vulnerabilities.The contribution statement reports both the prototype evaluation and security-capability analysis.

2 Related Work

Prior work spans hardware, software, and hardware-assisted isolation, but these approaches trade programmability, trusted-computing-base size, visibility, and complexity. TrustZone provides the isolated execution domains that LOOM uses for out-of-band monitoring, while prior monitors illustrate the visibility trade-off.

Isolation mechanisms: Hardware isolation mechanisms such as TPMs and HSMs support measured boot and secure key storage but have limited programmability and applicability.These limitations can make them unsuitable for broader monitoring tasks.
Isolation mechanisms: Microkernels, containers, and virtualization move services or entire operating systems outside the monolithic kernel but add system complexity and trusted components.The microkernel or hypervisor may itself become part of the trusted computing base and an attack target.
ARM TrustZone: ARMv8-A TrustZone divides execution into Secure and Non-secure states across processor cores, memory, and peripherals, managed by the Secure Monitor at EL3.Its software architecture includes a Rich Execution Environment and a Trusted Execution Environment.
Out-of-band monitoring: Out-of-band monitoring separates observation from the OS kernel, reducing the attack surface and improving resilience against kernel compromise.This separation supplies an independent monitoring domain rather than relying on the monitored kernel.
Monitoring approaches: Pure hardware monitors provide strong isolation but limited visibility, whereas VMI improves access to low-level state by monitoring from a privileged domain outside guest systems.The comparison frames the central trade-off between isolation and fine-grained kernel visibility.

3 Design

LOOM places monitoring in the TrustZone Secure World and reconstructs only selected kernel semantics supplied through a controlled Normal World channel. Its layered design adds targeted monitoring, address-translation caching, and threat-model boundaries for efficient rootkit detection.

3.1 Framework: LOOM separates Secure World monitoring from the Normal World kernel through a hardware-enforced boundary.This architecture prevents the normal-world kernel from directly controlling the monitoring component.
3.1 Framework: The framework follows principles of preserving isolation while accessing necessary information, reconstructing minimal semantics, responding promptly, and limiting overhead.These principles balance security, visibility, response time, and efficiency.
3.1 Framework: LOOM’s layered architecture assigns isolation and access to the platform layer, monitoring and response to the functional layer, and optimization to the auxiliary layer.The functional layer operates in the Secure World and monitors Normal World entities such as processes and system-call structures.
3.1 Framework: The Normal World driver exports selected kernel-object states through static shared memory, while the Secure World reconstructs task lists, module lists, file-system objects, and system-call contexts.Reconstruction avoids full kernel semantics and is performed without kernel cooperation.
3.1 Framework: An address-translation cache avoids repeated mappings for frequently accessed objects, reducing monitoring latency and overhead.The workflow polls at 10 Hz, with each monitoring task taking about 100 ms on average for address resolution and semantic reconstruction.
3.2 Threat Model: LOOM assumes trusted TrustZone hardware, secure boot, and Secure World software while focusing detection on security-critical objects linked to specific rootkit surfaces.It excludes physical hardware attacks, side channels, and resource-exhaustion denial-of-service attacks.

4 Implementation

LOOM implements out-of-band monitoring through physical address resolution and lightweight semantic reconstruction in the Secure World. It selectively reconstructs security-critical kernel attributes and combines semantic checks with dual-stage hazard prevention under TrustZone isolation.

Lightweight Out-of-Band Semantic Reconstruction: LOOM’s monitoring pipeline has two stages: hierarchical page-table traversal resolves physical addresses, then semantic reconstruction interprets selected kernel objects.The design emphasizes semantic reconstruction and reasoning under strict isolation rather than introducing a new address-translation mechanism.
Physical Address Resolution: Direct physical-address resolution mitigates page-table address-mapping deception and provides the prerequisite for subsequent semantic reconstruction.LOOM traverses configured page-table levels and safely accesses the resulting physical addresses from the Secure World.
Lightweight Semantic Reconstruction: LOOM bridges the semantic gap by reconstructing only bounded, security-relevant attributes instead of complete kernel data structures.The approach targets objects such as process descriptors, system call tables, and kernel modules, avoiding full structural traversal and redundant recovery.
Kernel-Independent Semantic Parsing Model: Boot-time metadata resolution avoids fixed kernel offsets and supports adaptive semantic descriptors across kernel instances, subject to relatively stable target-object organization.LOOM uses essential metadata such as System.map and resolves descriptor parameters during initialization.
Kernel-Independent Semantic Parsing Model: Semantic descriptors encode each monitored element using its logical structure, type, base address, and offset, enabling composable interpretation of raw memory.Logical structures represent access patterns such as linked lists, hierarchies, and indexed arrays, while type specifies the primitive data format and size.
Hazard Prevention: LOOM uses semantic consistency checks to trigger dual-stage prevention: hardware page locking protects fixed kernel pointers, while an adaptive agent mitigates flexible-footprint attacks.Pointer-hijacking attacks target objects such as sys_call_table and file_operations; dynamically loadable modules and hidden processes receive active, context-aware mitigation.

5 Evaluation

LOOM is evaluated for performance, communication overhead, semantic reconstruction cost, and capability-oriented trade-offs. The experiments report negligible Normal World overhead, much lower shared-memory communication cost than world switching, and substantial cache benefits for repeated address translation.

Experimental setup: LOOM was evaluated on an ARMv8-A Phytium D2000 platform with Linux in the Normal World and OP-TEE in the Secure World.FreeRTOS was ported into OP-TEE, with Secure World timer interrupts driving periodic monitoring independently of Normal World scheduling.
RQ1: Normal World performance: UnixBench reports 0.6% single-core and 0.8% multi-core overhead, both within the ±2% error margin.The results indicate negligible performance overhead for Normal World execution.
RQ2: Communication overhead: Shared-memory retrieval requires approximately 3.5 × 10^4 CPU cycles, compared with 7.6 × 10^8 cycles for an SMC round-trip.The comparison indicates that REE–TEE world switches dominate communication overhead.
RQ3: Semantic overhead reduction: Cache hits require 2.7 × 10^7 cycles, approximately 6.3× faster than cache misses at 1.7 × 10^8 cycles.Without caching, semantic reconstruction requires approximately 1.47 × 10^8 cycles per resolution, while address translation dominates semantic overhead.
RQ3: Semantic overhead reduction: Caching becomes beneficial above a 0.21 hit rate, and an 80% hit rate reduces average access latency by roughly 60%.The cache uses a hash table of size 2^n and computes indices with a bitwise AND operation.
Capability-oriented comparison: LOOM is compared with prior systems by isolation, semantic visibility, overhead, and trusted computing base or attack surface rather than direct numerical performance.The paper states that different execution models make direct numerical comparisons often not directly meaningful.

6 Security Capability Analysis

LOOM analyzes defensive capability through representative CVE cases involving privilege escalation, process hiding, and syscall hijacking. It reconstructs critical kernel-object semantics and checks integrity, references, lifecycle state, and memory protections to identify or mitigate these behaviors.

Scope and targets: The CVE analysis targets broader exploit classes, including privilege escalation, process hiding, and syscall hijacking, rather than a single exploit chain.The selected cases are used to assess semantic correctness of security-critical kernel states.
Monitored kernel objects: LOOM monitors init_task, init_cred, and sys_call_table as frequently targeted security-critical kernel objects.Their reconstructed semantics connect kernel-object monitoring to real-world vulnerability scenarios.
Process information: For init_task, LOOM checks PID, task state, executable name, and task-list pointers to identify process hiding, Netfilter exploitation, and process injection vulnerabilities.Task-list integrity and abnormal changes in the comm field support detection of these attack behaviors.
Process credentials: For init_cred, LOOM monitors credential pointers, reference counts, and user IDs to detect privilege escalation and track use-after-free conditions.Invalid references to init_cred support detection of CVE-2021-33909, while lifecycle and reference-count tracking addresses CVE-2022-2588 and CVE-2021-4154.
System call table: For sys_call_table, LOOM verifies syscall pointer integrity and page protection attributes to detect tampering, illegal writes, and rootkit behavior.Atomic hazard prevention is used to detect illegal writes to the syscall-table page in the CVE-2016-5195 example.

7 Conclusion

LOOM places lightweight out-of-band monitoring in the ARM TrustZone Secure World and selectively reconstructs critical kernel-object semantics. Its prototype and experiments report negligible performance overhead while combining isolation, semantic analysis, and dual-stage hazard prevention for secure kernel monitoring.

Conclusion: LOOM locates monitoring logic in the Secure World to remove the shared-fate dependency between protection mechanisms and a potentially compromised kernel.The architecture uses ARM TrustZone-based out-of-band monitoring for kernel security.
Conclusion: Selective semantic reconstruction captures critical kernel-object state while minimizing semantic overhead.The approach is paired with dual-stage hazard prevention for proactive sensitive-memory protection and rapid response to malicious kernel behavior.
Conclusion: Prototype implementation and experimental evaluation show negligible performance overhead for LOOM.The paper presents LOOM as an efficient and practical foundation for hardware–software co-designed secure kernel monitoring.