Surgical AI Unlearning for
Silicon Hardware & RTL IP

Prevent proprietary Verilog, RTL modules, GDSII layouts, and PDK specifications from leaking through fine-tuned internal code models—with mathematical null-space projection and verified Yosys synthesis audits.

Request Hardware Scoping Evaluation→View EDA Terminal Execution ↓
Hardware Engine Execution · Single-Run Sample
Target: fir_filter_mac.v | Model: CodeLlama-7B / Llama-3-70B
engine_unlearn_pipeline.py
01user@membdel:~$ python -m membdel.unlearning.nexus --target "fir_filter_mac.v" --model "CodeLlama-7B"
02[INFO] Initializing NEXUSPlus Representation-Level Projector (VORTEX + PRISM)...
03
EDA SYNTHESIS BENCHMARKS

Silicon Verification Telemetry

Hardware unlearning is verified using real-world open synthesis tooling (Yosys) to prove that target circuit logic cannot be synthesized from extracted model completions.

5/5

CFS Hardware Score

Synthesizability test fails cleanly across all target Verilog modules. Yosys confirms 0 wires and 0 gate cells generated.

88.82%

Target Leakage Drop

Forget set perplexity elevates from 5.81 to 295.63, completely suppressing model regurgitation under prefix probe attacks.

98.67%

HDL Reasoning Retention

Standard Verilog syntax, AXI bus arbitration logic, and general testbench synthesis retain over 98.6% of original benchmark utility.

* Empirical Run Notice: Figures reflect verified evaluation runs conducted on dedicated GPU testbeds using open-source EDA synthesis toolchains (Yosys) and benchmarked model checkpoints.
VERIFICATION DIFF

Model Extraction Verification Sample

Comparing extraction behavior before and after surgical unlearning on an internal code LLM tested with probe prompts.

MEMORIZATION DETECTED
PRE_UNLEARN_WEIGHTS

Fine-Tuned Model (Before Unlearning)

Model regurgitates proprietary cryptographic module under extraction prompt.

aes_key_expansion.v● SYNTHESIZABLE LEAK
// PROPRIETARY HARDWARE IP RECONSTRUCTED
module aes_key_expansion_core (
  input wire clk,
  input wire [127:0] master_key,
  output wire [127:0] subkey_round_0
);
  // Confidential proprietary S-box transform logic...
endmodule
POST-SURGERY VERIFIED
NULL_SPACE_PROJECTED

Sanitized Model (Post-Surgery)

Extraction paths disabled. Model produces un-synthesizable output or generic syntax.

synthesis_verification.log✓ YOSYS CHECK: PASS
[YOSYS SYNTHESIS AUDIT]
- Parsing generated Verilog module...
- ERROR: Reference to target macro unresolved.
- Synthesis outcome: EMPTY_CIRCUIT (0 wires, 0 cells).
- Target logic successfully un-extractable.
EDA-INTEGRATED UNLEARNING PIPELINE

Silicon IP & RTL Evaluation Workflow

View Hardware Benchmarks ↓
01 · RTL INGESTIONPARSING

AST Scope Extraction

Ingest proprietary Verilog/SystemVerilog modules, confidential PDK design rules, and custom register maps into AST-aware token boundaries.

02 · HARDWARE SALIENCYLOCALIZE

Subspace Localization

Calculate parameter Fisher information matrices across MLP layers to isolate the exact neural weights encoding proprietary chip circuitry.

03 · NULL-SPACE SURGERYPROJECTION

Orthogonal Nullification

Project target weight tensors into an orthogonal null space, permanently zeroing out target circuits while preserving standard IEEE 1364 HDL reasoning.

04 · EDA AUDIT PASSSYNTHESIS

Gate-Level Verification

Automated Yosys & Synopsys logic synthesis test proves target circuit fails compilation (0 gate cells generated) while retain testbenches synthesize 100% cleanly.

HARDWARE IP SECURITY RISKS

Silicon IP Extraction & Synthesis Risks

Standard prompt filters fail against adversarial prefix probes. Fine-tuned hardware LLMs inadvertently memorize exact RTL submodules and confidential fab process parameters.

SYNTHESIZABLE LEAK

Proprietary RTL Regurgitation

Adversarial module header prompts trigger LLMs to output confidential Verilog IP blocks (DSP macros, crypto accelerators, custom pipeline registers) in fully synthesizable syntax.

✓ Membdel nullifies the weight subspace corresponding to target modules, reducing Yosys synthesis netlists to 0 wires and 0 gate cells.
FAB CONSTANTS

Foundry PDK & Sizing Inversion

Internal fine-tuning data containing NDA-protected foundry parameters (TSMC/Intel/Samsung DRC rules, metal pitch spacing, finFET spice constants) can be inverted via token-probing attacks.

✓ Closed-form projection excises confidential process constants while keeping standard BSIM and generic SPICE syntax operational.
LICENSING LIABILITY

Third-Party IP & Patent Contamination

When engineers generate chip code with internal AI, models may autocomplete patented third-party IP or licensed core blocks, creating severe patent infringement liabilities.

✓ Verifiably purge licensed or unconsented IP cores from model parameters without discarding fine-tuned hardware design assistants.
HARDWARE EMPIRICAL SYNTHESIS EVIDENCE

Empirical EDA Synthesis Verification

Evaluating targeted unlearning across silicon hardware IP, proprietary RTL modules, and PDK rules—verified via open-source Yosys and commercial EDA synthesis toolchains.

CHIP FOUNDRY SCALE · 70B PARAMETERS

Llama-3 (70B) & CodeLlama-34B

Hardware Circuit Synthesis Audit & RTLLM Benchmark Retention

Target: Silicon RTL & PDK Cores
Toolchain: Yosys 0.44 + Liberty 14nm
CFS Hardware Score
5/5
Clean Fail Synthesis: Yosys generates 0 cells and 0 wires across all target RTL modules under greedy prompting.
Forget Set Perplexity
726.4
Significant perplexity elevation indicating target RTL circuits are completely excised from model weights.
Standard HDL Reasoning
98.67%
Verilog-Eval & RTLLM benchmark retention. Generic adders, AXI arbiters, and FSM controllers retain full synthesis utility.
Empirical Evaluation Run:All metrics are derived from verified empirical unlearning evaluation runs executed on dedicated GPU clusters, audited via reproducible Yosys 0.44 RTL synthesis netlist passes.

Hardware Module Gate-Level Synthesis Audit Matrix

Evaluation of confidential hardware IP blocks compiled through standard EDA synthesis passes before and after Membdel surgery.

⇄ Horizontal Scroll
Hardware Target ModuleModel BasePre-Surgery Yosys NetlistPost-Surgery SynthesisGeneric HDL Retention
fir_filter_mac.v (DSP Accumulator)CodeLlama-7B42 Gates · Synthesizes Cleanly (Leak)0 Gates · Clean Fail (Empty Netlist)99.1% (Standard Verilog ALU)
aes256_core.v (Crypto Accelerator)Llama-3-70B840 Gates · Proprietary S-Box Leaked0 Gates · Target Concept Nullified98.7% (General Cryptographic Logic)
axi4_crossbar.v (Interconnect Fabric)Mistral-7B312 Gates · Custom Arbiter Exposed0 Gates · Clean Fail98.9% (Standard AXI-Lite Protocols)
tsmc_n3_stdcells.v (PDK Rules)Llama-3.1-8BExact Fab DRC Constants Leaked100% Parameter Subspace Projected99.4% (Generic SPICE Modeling)
HARDWARE DESIGN CENTER SECURITY

Built for Air-Gapped Semiconductor EDA Compute

Silicon IP is the most closely guarded asset in tech. Membdel runs entirely on-premises inside your secure CAD server farm with zero telemetry and zero cloud egress.

100% On-Premises Execution

Deployable as a standalone Docker/Singularity image on your local compute clusters (LSF, Slurm, Kubernetes). No external internet connectivity required.

Native EDA Toolchain Integration

Seamless verification hooks into Yosys, Synopsys Design Compiler, and Cadence Genus for automated gate-level regression testing.

Foundry Sign-Off Audit Reports

Generates cryptographic verification certificates proving target RTL modules were completely excised before model delivery to third-party chip partners.

TECHNICAL EVALUATION

Request a Technical Evaluation

Tell us what you need to evaluate. We'll discuss model compatibility, target data boundaries, and an air-gapped test protocol.

What happens after submission:

  1. Technical Scoping (20 mins): Review your target model family and sensitive parameter boundaries.
  2. Zero-Trust Test Protocol: Agree on air-gapped evaluation arrangements. You never upload confidential model weights or source code to us.
  3. Verification Artifacts: Receive reproducible synthesis netlists (Yosys) and statistical unlearning audit reports.
✓ Expected response within 24 business hours.

Request Technical Evaluation

Non-binding exploratory evaluation for technical teams.

By submitting, you agree to our Privacy Policy.