Zero-knowledge proofs (ZKPs) are increasingly important for privacy-preserving and encrypted computation, including federated learning, verifiable inference, and secure data analytics. However, most existing ZKP accelerators target ZKP protocols based on arithmetic circuits over large prime fields. These are often rendered inefficient for Boolean-heavy computations, where native operations are bit-level AND and XOR gates rather than large-field additions and multiplications. As encrypted-AI systems increasingly incorporate quantized, Boolean, and cryptographic components, this Boolean-to-arithmetic mismatch becomes a key obstacle to efficient proof generation. This paper explores Vector Oblivious Linear Evaluation-in-the-Head (VOLEitH) as an alternative protocol substrate for publicly verifiable Boolean ZKPs. VOLEitH transforms efficient designated-verifier VOLE-based ZK protocols into non-interactive, publicly verifiable, post-quantum-secure proofs. We present VOLE-ocity, an FPGA-based accelerator for VOLEitH proof generation. VOLE-ocity decomposes the prover into hardware-optimized modules for GGM-tree vector commitments, streaming VOLE conversion, and gate-by-gate Boolean circuit evaluation. Our prototype, implemented with HLS on a Xilinx Virtex UltraScale+ FPGA, achieves up to 28× speedup over highly optimized CPU baselines on SHA-256 and Keccak-f circuits. Finally, we discuss how hardware-accelerated Boolean ZK can serve as a foundation for future VOLE-based machine-learning proof systems, particularly for quantized and Boolean neural-network workloads.

Encrypted AI using fully homomorphic encryption (FHE) provides strong privacy guarantees; but its slow performance has limited practical deployment. Recent works proposed ASICs to accelerate FHE, but require expensive advanced manufacturing processes that constrain their accessibility. GPUs are a far more accessible platform, but achieving ASIC-level performance using GPUs has remained elusive. Furthermore, state-of-the-art approaches primarily focus on small models that fit comfortably within a single device. Supporting large models such as LLMs in FHE introduces a dramatic increase in computational complexity that requires optimized GPU kernels, along with managing terabyte-scale memory footprints that far exceed the capacity of a single GPU. This paper presents Cerium, a multi-GPU framework for FHE inference on large models. Cerium integrates a domain-specific language, an optimizing compiler, and a runtime system to automatically generate high-performance GPU kernels, manage terabyte-scale memory footprints, and parallelize computation across multiple GPUs. It introduces new IR constructs, compiler passes, sparse polynomial representations, memory-efficient data layouts, and communication-aware parallelization techniques that together enable encrypted inference for models ranging from small CNNs to Llama3-8B. We build Cerium on NVIDIA GPUs and demonstrate significant performance gains. For small models, Cerium outperforms expert-written hand-optimized GPU libraries by up to 2.25 times. Cerium achieves performance competitive with state-of-the-art FHE ASICs, outright matching prior FHE ASIC CraterLake. It is the first GPU system to execute bootstrapping in under 10 milliseconds, achieving 7.5 milliseconds, and is the first to demonstrate encrypted inference for BERT-Base and Llama3-8B in 8 seconds and 43 seconds, respectively.

Zero-Knowledge Proofs (ZKPs) enable privacy-preserving verification but remain difficult to deploy due to high proof-generation costs. Recent work has explored hardware and GPU acceleration across diverse ZKP schemes, each with different trade-offs in overhead, proof size, verifier cost, setup, and trust. This work targets a key bottleneck in modern ZKP systems: supporting complex, high-degree custom gates through the SumCheck protocol. We propose a programmable accelerator for arbitrary SumCheck-based gates, achieving over 1000× geomean speedup over CPU SumChecks across multiple gate types. Integrated into zkPHIRE, a full-system programmable accelerator for HyperPlonk, the design achieves 1486× geomean speedup over CPU and 11.87× 11.87× over prior state-of-the-art at iso-area, while scaling beyond 2^30 constraints and maintaining compact 4-5 KB proofs.

Zero-knowledge proofs of neural-network inference (zkML) can certify model execution, but proving cost remains a central obstacle, especially for transformer-style models. FLOP counts are an incomplete proxy: field additions and multiplications are relatively natural for arithmetic proof systems, while comparisons, lookups, divisions, normalization, and transcendental functions introduce proof-unfriendly overheads. We present ZKTrace, a lightweight backend-agnostic profiler that traces small PyTorch models with torch.fx and reports both a FLOP/MAC baseline and an operator-aware proof-cost score. We validate on six approximately FLOP/MAC-matched model pairs using ezkl. In five of six pairs the higher-scored member proves slower; across all twelve models the proof-aware score ranks measured proving time at Spearman 0.83 versus 0.59 for the baseline. The result is preliminary: ZKTrace is an early-stage design tool, not a proving system or exact prover-time predictor.

Fully Homomorphic Encryption (FHE) allows computation on encrypted data without decrypting it. In theory, FHE makes privacy-preserving machine learning possible. In practice, however, it remains extremely slow for real workloads. A major source of slowdown is bootstrap operations; in CKKS, a popular FHE scheme for tensor workloads, the slowdown is compounded by scale management and rescale operations. Existing FHE compilers miss crucial optimization opportunities because they overlook a key property of CKKS programs: bootstrap and rescale placement are fundamentally coupled through the level budget. We present ORBIT, an FHE compiler that jointly optimizes bootstrap and rescale placement through a novel Integer Linear Programming (ILP) formulation that reasons about both ciphertext level and scale constraints. To make this formulation tractable for structured tensor workloads, we introduce three techniques that reduce ILP complexity while preserving optimality. Across five convolutional neural networks and multiple cryptographic parameter configurations, ORBIT achieves a geometric mean speedup of up to 19% over DACAPO, 73% over ORION, and 52% over RESBM, keeps compilation under 6 minutes, and retains model accuracy within 0.3% of plaintext execution.