In this work we study bounded collusion protocols (BCPs) recently introduced in the context of secret sharing by Kumar, Meka, and Sahai (FOCS 2019). These are multi-party communication protocols on $n$ parties where in each round a subset of $p$-parties (the collusion bound) collude together and write a function of their inputs on a public blackboard.

BCPs interpolate elegantly between the well-studied number-in-hand (NIH) model ($p=1$) and the number-on-forehead (NOF) model ($p=n-1$). Motivated by questions in communication complexity, secret sharing, and pseudorandomness we investigate BCPs more thoroughly answering several questions about them.

* We prove a polynomial (in the input-length) lower bound for an explicit function against BCPs where any constant fraction of players can collude. Previously, non-trivial lower bounds were only known when the collusion bound was at most logarithmic in the input-length (owing to bottlenecks in NOF lower bounds).

* For all $t \leq n$, we construct efficient $t$-out-of-$n$ secret sharing schemes where the secret remains hidden even given the transcript of a BCP with collusion bound $O(t/\log t)$. Prior work could only handle collusions of size $O(\log n)$. Along the way, we construct leakage-resilient schemes against disjoint and adaptive leakage, resolving a question asked by Goyal and Kumar (STOC 2018).

* An explicit $n$-source cylinder intersection extractor whose output is close to uniform even when given the transcript of a BCP with a constant fraction of parties colluding. The min-entropy rate we require is $0.3$ (independent of collusion bound $p \ll n$).

Our results rely on a new class of exponential sums that interpolate between the ones considered in additive combinatorics by Bourgain (Geometric and Functional Analysis 2009) and Petridis and Shparlinski (Journal d'Analyse Mathématique 2019).