PhD student in computer systems at UCAS / IIE, working under
Prof. Rui Hou and
Prof. Wenhao Wang.
My work spans hardware ISA, firmware, Linux kernel internals, and userspace —
independently built. Past projects: NestedSGX (built nested enclaves within confidential VMs;
NDSS'25), a solo-built confidential container system (custom ISA + kernel
critical-path patching + firmware enforcement; now being ported to the BOOM
RISC-V core and FPGA platforms), and an LLM KV-cache side-channel
attack (led GPU noise mitigation, attack design, and implementation;
99% TPR, TIFS'25).
Research interests: OS-level isolation, virtualization, sandboxing, TEEs,
scalable confidential computing, and LLM agent infrastructure (container-based
sandboxing, KV cache security, privacy-preserving inference).
CS self-study.
Seeking internship opportunities in agent systems, agent
container/sandboxing, and KV cache infrastructure.
I am currently seeking relevant internship opportunities. If any of this
resonates, I would be glad to connect — thank you! 🙏
我是宋林轲,在读的 persistent head damage 一名,目前的状态是:I need a doctor。
兴趣主要是兴趣使然,玩过的东西不少,总体上样样通、样样松,只有弹琴和跑步还算长期在线。
喜欢听曲子、扒谱和随手即兴,没有绝对音感,但相对音感够用,听过的曲子大多能在琴上摸出来;
手上功夫大概在 Animenz 入门曲附近。每周跑三次左右,每次长跑 5–10 公里,也是个驿传粉丝。
Problem: Existing confidential containers repurpose hardware mechanisms not designed for containers, sacrificing native-ness. Arm CCA treats containers as TEE workloads with large contiguous memory — no primitives purpose-built for native confidential containers.
Approach: Reused RISC-V PMP with bitmap + MMU modifications to protect page-table pages and data pages without fragmenting Linux's native memory allocation — containers stay ordinary processes.
vs. Existing: No virtualization/confidential-computing hardware dependency (unlike Kata/gVisor). Introduces a new hardware abstraction layer for containers; native Linux allocation with targeted hardening, no invasive kernel patches.
Security: Container lifecycle protection integrated into the full address-allocation pipeline.
Performance: Near-zero overhead vs. native Docker on compute-intensive tasks; under 2× slowdown on memory-intensive tasks — significantly better than microVM approaches.
Hardware porting (in progress): Porting NACRE beyond its QEMU model onto real hardware platforms — the BOOM out-of-order RISC-V core (via Chipyard) and FPGA prototyping (FireSim flow), parameterizing the physical layout (OpenSBI/Linux/Agent/metadata/PTP regions) for FPGA DRAM configurations.
Porting work: Jointly discussed and finalized the hardware design with Yifan Bai, who proposed the Agent Mode programming model and implemented it based on the RISC-V H extension; I authored the QEMU↔RTL semantic-alignment design that freezes the hardware contract for the port — custom CSRs (nacc_state, nacc_sagent/nacc_eagent, twin_entry, trampoline), acall/aret custom instructions, PTW-time page-table checks, the per-PFN 2-bit metadata bitmap, and TLB revocation on tag change.
My role: Sole developer spanning the entire stack — custom hardware ISA primitives, Linux kernel critical-path modifications (page allocation, RSS accounting, page-fault handling — preserving existing metadata correctness), OpenSBI firmware integration, ecall interface and protocol design. The Agent Mode programming model for the hardware port was proposed by Yifan Bai.
LLM Side-Channel Attack on KV Cache2024.6 – 2025.4
Problem: KV cache reuse in LLM serving systems creates a timing side channel — TTFT reductions reveal victim prompt overlap. Found exploitable in SGLang (prefix sharing) and GPT-Cache (semantic similarity).
Approach: In SGLang, a single shared token with Llama 3.1 8B/70B yields measurable latency drop. Designed token-by-token prompt recovery; overcame GPU voltage/frequency noise with a countermeasure, achieving 99% TPR.
GPT-Cache Attack: Discovered a distinct vector — semantically similar queries with identical sensitive info trigger TTFT speedups via similarity-based cache matching.
Defense: Proposed coarse-grained token-sharing defense that expands the attacker's guessing space.
Impact: One of the earliest two teams to report KV-cache privacy risks to SGLang (the other: ByteDance Security Research, same week). Presented at SGLang biweekly meeting (Oct. 19, 2024).
Publication: Submitted to USENIX Security'25, ACM CCS'25; accepted at TIFS'25 (CCF-A journal).
NestedSGX: Nested Enclaves in Confidential VMs2023.7 – 2024.5
Role: co-lead | Rust · C · Python · Linux Kernel Module · Qemu · AMD SEV-SNP |
code |
paper
Problem: Confidential VMs face a large TCB in the guest OS. No existing mechanism to establish trusted enclaves inside a CVM while keeping the guest OS out of the TCB.
Approach: Leveraged AMD SEV-SNP VMPL to introduce a lightweight hypervisor within the CVM, de-privileging the guest OS — even a compromised kernel cannot access enclave memory.
Compatibility: Built trusted enclave runtime atop Occlum and Intel SGX SDK, compatible with existing SGX ecosystem — unmodified SGX apps run inside the nested enclave.
Engineering: Modified Linux kernel drivers; wrote in-VM hypervisor in Rust (7,000+ LoC — page faults, error paths, custom trampoline for cross-privilege transitions).
Recognition:2 AE badges; invited by Asterinas community for online seminar. Submitted to ASPLOS'24, ACM CCS'24; accepted at NDSS'25 (CCF-A conference).
Publications
The Early Bird Catches the Leak: Unveiling Timing Side Channels in LLM Serving Systems