papers/hanzo-aci/sections/10-discussion.tex

\section{Discussion}

\label{sec:discussion}

\subsection{Privacy}

ACI currently requires workers to see job inputs in plaintext (within the TEE). Future work will explore:
\begin{itemize}[leftmargin=1.1em]
    \item Encrypted inference using homomorphic encryption (currently too slow for practical use).
    \item Secure multi-party computation for splitting jobs across multiple workers.
    \item Differential privacy for training workloads.
\end{itemize}

\subsection{Limitations}

\begin{enumerate}[leftmargin=1.4em]
    \item \textbf{TEE dependency}: PoAI's attestation layer relies on TEE hardware, which has known side-channel vulnerabilities. Statistical verification mitigates but does not eliminate this risk.
    \item \textbf{Data transfer}: Large model weights and datasets must be transferred to workers, creating bandwidth bottlenecks for distributed training.
    \item \textbf{Non-determinism}: Floating-point non-determinism across GPU architectures complicates exact verification.
    \item \textbf{Regulatory}: Decentralized compute markets face regulatory uncertainty in some jurisdictions.
\end{enumerate}

\subsection{Future Work}

\begin{itemize}[leftmargin=1.1em]
    \item \textbf{ZK-PoAI}: Hybrid verification combining ZK proofs for deterministic operations with statistical sampling for stochastic operations.
    \item \textbf{Cross-chain bridges}: Enable ACI to serve as a compute layer for other blockchains.
    \item \textbf{Federated training}: Extend PoAI to verify federated learning contributions.
    \item \textbf{Hardware oracle}: Integrate hardware benchmarking into worker registration to validate claimed capabilities.
\end{itemize}