Prerequisites & Notation

Before You Begin

Chapter 11 carries the third CommIT-group contribution of Part III: the ByzSecAgg protocol that simultaneously handles privacy, dropouts, and Byzantine adversaries. The chapter composes machinery from multiple earlier chapters: Shamir/ramp sharing (Ch 3), polynomial codes (Ch 5), gradient coding (Ch 6), and Bonawitz secure aggregation (Ch 10). Readers comfortable with all four will find the construction modular and natural.

  • Shamir + ramp secret sharing (Chapter 3)(Review ch03)

    Self-check: State the share-size advantage of ramp sharing (gg-fold smaller) and the leakage tradeoff.

  • Polynomial codes for matrix multiplication (Chapter 5)(Review ch05)

    Self-check: Why does the polynomial code achieve recovery threshold K=pqK = pq?

  • Gradient coding (Chapter 6)(Review ch06)

    Self-check: Recall how gradient coding tolerates ss stragglers at per-worker storage (s+1)/N(s+1)/N.

  • Bonawitz secure aggregation (Chapter 10)(Review ch10)

    Self-check: Restate the threat model and Bonawitz's O(n2)O(n^2) cost.

  • Cryptographic vector commitments (Merkle trees, Pedersen)

    Self-check: Can you sketch a Merkle tree for committing to a vector and prove a single-element opening in O(log⁑n)O(\log n)?

Notation for This Chapter

Byzantine-resilient secure aggregation extends Chapter 10's notation with the Byzantine threshold BB (number of malicious users), the ramp-sharing parameters (t1,t2)(t_1, t_2), and a vector-commitment scheme. The ramp width g=t2βˆ’t1g = t_2 - t_1 controls the share-size savings.

SymbolMeaningIntroduced
nnNumber of userss01
BBMaximum number of Byzantine (malicious) userss01
TTPrivacy threshold (colluding honest-but-curious users)s01
gk\mathbf{g}_kUser kk's gradient (honest) or arbitrary vector (Byzantine)s01
ggRamp width = t2βˆ’t1t_2 - t_1 (Chapter 3)s02
Ck\mathcal{C}_kVector commitment to user kk's gradients02
Ο€k\pi_kOpening / proof of Ck\mathcal{C}_ks02
dkjd_{kj}Coded distance score between users kk and jj (used for outlier detection)s03
ddModel dimensionalitys01