arXiv subjects

Allison Bishop

Publications and source records attributed to Allison Bishop.

6 recordsLinked to original sources

arXiv2026

Context Compaction Theory

Large Language Models (LLMs) have a bounded context window. The context window is the maximum input size an LLM can consume for a single inference. AI agents rely on a process called context compaction to fit their state within the context window when calling an LLM. Despite its ubiquity, context compaction has received essentially no formal analysis. In this paper, we initiate a formal study of context compaction. We first introduce a framework consisting of two games that capture the two algorithmic strategies for context compaction used by contemporary AI agents in practice. The Context Selection Game models context compaction algorithms that select a subset of an agent's accumulated state to retain. The Context Generation Game models context compaction algorithms that summarize an agent's state by an arbitrary message of bounded length. We then prove an equivalence between the Context Generation Game and one-way communication complexity. The minimum context compaction budget for answering a set of queries within a target error is equal to the one-way communication complexity of the induced communication problem at the same error. Known bounds from communication complexity therefore transfer directly to context compaction. We also show that the Context Selection Game corresponds to a restricted class of one-way communication protocols. Any gap between selection and generation is therefore a gap between two classes of communication protocols. We prove that there exists a set of queries for which generation needs strictly less budget than selection. The equivalence between the Context Generation Game and one-way communication also lets us measure how well a deployed context compaction algorithm performs on a query relative to the optimal strategy. As an example, we present a case study that evaluates Anthropic's context compaction endpoint on set membership queries.

cs.DS↗

arXiv2024

Adversary Resilient Learned Bloom Filters

A learned Bloom filter (LBF) combines a classical Bloom filter (CBF) with a learning model to reduce the amount of memory needed to represent a given set while achieving a target false positive rate (FPR). Provable security against adaptive adversaries that advertently attempt to increase FPR has been studied for CBFs, but not for LBFs. In this paper, we close this gap and show how to achieve adaptive security for LBFs. In particular, we define several adaptive security notions capturing varying degrees of adversarial control, including full and partial adaptivity, in addition to LBF extensions of existing adversarial models for CBFs, including the Always-Bet and Bet-or-Pass notions. We propose two secure LBF constructions, PRP-LBF and Cuckoo-LBF, and formally prove their security under these models assuming the existence of one-way functions. Based on our analysis and use case evaluations, our constructions achieve strong security guarantees while maintaining competitive FPR and memory overhead.

cs.CR↗

arXiv2020

A cryptographic approach to black box adversarial machine learning

We propose a new randomized ensemble technique with a provable security guarantee against black-box transfer attacks. Our proof constructs a new security problem for random binary classifiers which is easier to empirically verify and a reduction from the security of this new model to the security of the ensemble classifier. We provide experimental evidence of the security of our random binary classifiers, as well as empirical results of the adversarial accuracy of the overall ensemble to black-box attacks. Our construction crucially leverages hidden randomness in the multiclass-to-binary reduction.

cs.LG↗

arXiv2019

MicroCash: Practical Concurrent Processing of Micropayments

Micropayments are increasingly being adopted by a large number of applications. However, processing micropayments individually can be expensive, with transaction fees exceeding the payment value itself. By aggregating these small transactions into a few larger ones, and using cryptocurrencies, today's decentralized probabilistic micropayment schemes can reduce these fees. Unfortunately, existing solutions force micropayments to be issued sequentially, thus to support fast issuance rates a customer needs to create a large number of escrows, which bloats the blockchain. Moreover, these schemes incur a large computation and bandwidth overhead, which limit their applicability in large-scale systems. In this paper, we propose MicroCash, the first decentralized probabilistic framework that supports concurrent micropayments. MicroCash introduces a novel escrow setup that enables a customer to concurrently issue payment tickets at a fast rate using a single escrow. MicroCash is also cost effective because it allows for ticket exchange using only one round of communication, and it aggregates the micropayments using a lottery protocol that requires only secure hashing. Our experiments show that MicroCash can process thousands of tickets per second, which is around 1.7-4.2x times the rate of a state-of-the-art sequential micropayment system. Moreover, MicroCash supports any ticket issue rate over any period using only one escrow, while the sequential scheme would need more than 1000 escrows per second to permit high rates. This enables our system to further reduce transaction fees and data on the blockchain by around 50%.

cs.CR↗

arXiv2019

ABC: A Cryptocurrency-Focused Threat Modeling Framework

Cryptocurrencies are an emerging economic force, but there are concerns about their security. This is due, in part, to complex collusion cases and new threat vectors that could be missed by conventional security assessment strategies. To address these issues, we propose ABC, an Asset-Based Cryptocurrency-focused threat modeling framework capable of identifying such risks. ABC's key innovation is the use of collusion matrices. A collusion matrix forces a threat model to cover a large space of threat cases while simultaneously manages this process to prevent it from being overly complex. Moreover, ABC derives a system-specific threat categories that account for the financial aspects and the new asset types that cryptocurrencies introduce. We demonstrate that ABC is effective by conducting a user study and by presenting real-world use cases. The user study showed that around 71$\%$ of those who used ABC were able to identify financial security threats, as compared to only 13$\%$ of participants who used the popular framework STRIDE. The use cases further attest to the usefulness of ABC's tools for both cryptocurrency-based systems, as well as a cloud native security technology. This shows the potential of ABC as an effective security assessment technique for various types of large-scale distributed systems.

cs.CR↗

arXiv2019

CAPnet: A Defense Against Cache Accounting Attacks on Content Distribution Networks

Peer-assisted content distribution networks(CDNs) have emerged to improve performance and reduce deployment costs of traditional, infrastructure-based content delivery networks. This is done by employing peer-to-peer data transfers to supplement the resources of the network infrastructure. However, these hybrid systems are vulnerable to accounting attacks in which the peers, or caches, collude with clients in order to report that content was transferred when it was not. This is a particular issue in systems that incentivize cache participation, because malicious caches may collect rewards from the content publishers operating the CDN without doing any useful work. In this paper, we introduce CAPnet, the first technique that lets untrusted caches join a peer-assisted CDN while providing a bound on the effectiveness of accounting attacks. At its heart is a lightweight cache accountability puzzle that clients must solve before caches are given credit. This puzzle requires colocating the data a client has requested, so its solution confirms that the content (or at least an amount of data within a pre-configured bound) has actually been retrieved. We analyze the security and overhead of our scheme in realistic scenarios. The results show that a modest client machine using a single core can solve puzzles at a rate sufficient to simultaneously watch dozens of 1080p videos. The technique is designed to be even more scalable on the server side. In our experiments, one core of a single low-end machine is able to generate puzzles for 4.26 Tbps of bandwidth - enabling 870,000 clients to concurrently view the same 1080p video. This demonstrates that our scheme can ensure cache accountability without degrading system productivity.

cs.CR↗