Anshika Agarwal

Caleb Ziems, Dora Zhao, Rose E. Wang et al.

Large Language Models (LLMs) are increasingly shaping the private and professional lives of users, with numerous applications in business, education, finance, healthcare, law, and science. With this rise in global influence comes greater urgency to build, evaluate, and deploy these systems in a manner that prioritizes not only technical capabilities but also human priorities. This work presents a framework for developing Human-Centered Large Language Models (HCLLMs), which integrates perspectives from Natural Language Processing (NLP), Human-Computer Interaction (HCI), and responsible AI. Considering the ethics, economics, and technical objectives of language modeling, we argue that model developers need to address human concerns, preferences, values, and goals, not only during a cursory post-training stage, but rather with rigor and care at every stage of the pipeline. This paper offers human-centered insights and recommendations for developers at each stage, from system design to data sourcing, model training, evaluation, and responsible deployment. Then we conclude with a case study, applying these insights to understand the future of work with HCLLMs.

Devashish Gosain, Anshika Agarwal, Sahil Shekhawat et al.

This paper presents a study of the Internet infrastructure in India from the point of view of censorship. First, we show that the current state of affairs---where each ISP implements its own content filters (nominally as per a governmental blacklist)---results in dramatic differences in the censorship experienced by customers. In practice, a well-informed Indian citizen can escape censorship through a judicious choice of service provider. We then consider the question of whether India might potentially follow the Chinese model and institute a single, government-controlled filter. This would not be difficult, as the Indian Internet is quite centralized already. A few "key" ASes (approx 1% of Indian ASes) collectively intercept approx 95% of paths to the censored sites we sample in our study, and also to all publicly-visible DNS servers. 5,000 routers spanning these key ASes would suffice to carry out IP or DNS filtering for the entire country; approx 70% of these routers belong to only two private ISPs. If the government is willing to employ more powerful measures, such as an IP Prefix Hijacking attack, any one of several key ASes can censor traffic for nearly all Indian users. Finally, we demonstrate that such federated censorship by India would cause substantial collateral damage to non-Indian ASes whose traffic passes through Indian cyberspace (which do not legally come under Indian jurisdiction at all).

Devashish Gosain, Anshika Agarwal, Sambuddho Chakravarty et al.

Decoy Routing, the use of routers (rather than end hosts) as proxies, is a new direction in anti-censorship research. Decoy Routers (DRs), placed in Autonomous Systems, proxy traffic from users; so the adversary, e.g., a censorious government, attempts to avoid them. It is quite difficult to place DRs so the adversary cannot route around them for example, we need the cooperation of 850 ASes to contain China alone. In this paper, we consider a different approach. We begin by noting that DRs need not intercept all the network paths from a country, just those leading to Overt Destinations, i.e., unfiltered websites hosted outside the country (usually popular ones, so that client traffic to the OD does not make the censor suspicious. Our first question is; How many ASes are required for installing DRs to intercept a large fraction of paths from, e.g., China to the top n websites (as per Alexa)? How does this number grow with n? Few ASes (approx. 30) intercept over 90% of paths to the top n sites, for n = 10, 20...200. Our first contribution is to demonstrate with real paths that the number of ASes required for a world-wide DR framework is small (approx. 30). Further, censor nations attempts to filter traffic along the paths transiting these 30 ASes will not only block their own citizens, but others residing in foreign ASes. Our second contribution in this paper is to consider the details of DR placement: not just in which ASes DRs should be placed to intercept traffic, but exactly where in each AS. We find that even with our small number of ASes, we still need a total of about 11,700 DRs.We conclude that, even though a DR system involves far fewer ASes than previously thought, it is still a major undertaking. For example, the current routers cost over 10.3 billion USD, so if DR at line speed requires all new hardware, the cost alone would make such a project unfeasible for most actors.