Scholarlink

<p>The emergence of multi-modal models such as Apple’s MM1 signifies a transition towards on-device artificial intelligence, diminishing dependence on cloud inference. However, quantum developments render classical cryptography vulnerable to data breach. We present QSAFE-MM1, a quantum-resilient security architecture that incorporates Federated Learning (FL), Fully Homomorphic Encryption (FHE), and lattice-based cryptography to enhance MM1’s security. Federated Learning (FL) facilitates decentralised training without the transmission of raw data, so safeguarding user privacy and attaining 94% processing efficiency, 1020&#xa0;J energy consumption, 7% per hour battery depletion, and a thermal increase of + 4&#xa0;°C. Fully Homomorphic Encryption (FHE) facilitates encrypted inference, preventing data breaches while processing; yet, it results in an 81% efficiency reduction, consumes 1600&#xa0;J, causes a 13% per hour energy drain, and increases temperature by 7&#xa0;°C. The complete QSAFE-MM1 stack (FL + FHE + DP) achieves 79% efficiency, with 1700&#xa0;J, 14%/hr, and + 8&#xa0;°C, indicating secure-performance trade-offs. Quantum resistance is attained by NIST-compliant lattice-based methods that are impervious to Shor’s algorithm, and asymmetric masking eliminates personally identifiable information during training. Empirical assessment verifies that QSAFE-MM1 maintains model accuracy (± 1.2% variance) and latency (&lt; 9% overhead) while ensuring post-quantum security. QSAFE-MM1 establishes a new standard for mobile AI security, harmonising quantum safety, user privacy, and performance under strict resource limitations, thereby presenting MM1 as a frontrunner in secure, on-device intelligence.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantum resilient security framework for privacy preserving AI in Apple MM1 on device architecture

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月3日

查看全文
  • Nauman Umer,
  • Miaolei Deng,
  • Yuhong Zhang,
  • Miao Zhang,
  • Sheheryar Khan

摘要

The emergence of multi-modal models such as Apple’s MM1 signifies a transition towards on-device artificial intelligence, diminishing dependence on cloud inference. However, quantum developments render classical cryptography vulnerable to data breach. We present QSAFE-MM1, a quantum-resilient security architecture that incorporates Federated Learning (FL), Fully Homomorphic Encryption (FHE), and lattice-based cryptography to enhance MM1’s security. Federated Learning (FL) facilitates decentralised training without the transmission of raw data, so safeguarding user privacy and attaining 94% processing efficiency, 1020 J energy consumption, 7% per hour battery depletion, and a thermal increase of + 4 °C. Fully Homomorphic Encryption (FHE) facilitates encrypted inference, preventing data breaches while processing; yet, it results in an 81% efficiency reduction, consumes 1600 J, causes a 13% per hour energy drain, and increases temperature by 7 °C. The complete QSAFE-MM1 stack (FL + FHE + DP) achieves 79% efficiency, with 1700 J, 14%/hr, and + 8 °C, indicating secure-performance trade-offs. Quantum resistance is attained by NIST-compliant lattice-based methods that are impervious to Shor’s algorithm, and asymmetric masking eliminates personally identifiable information during training. Empirical assessment verifies that QSAFE-MM1 maintains model accuracy (± 1.2% variance) and latency (< 9% overhead) while ensuring post-quantum security. QSAFE-MM1 establishes a new standard for mobile AI security, harmonising quantum safety, user privacy, and performance under strict resource limitations, thereby presenting MM1 as a frontrunner in secure, on-device intelligence.