Madhusudan's homepage

Research Interests

Software verification and formal methods; Interpretable, Robust, Trustworthy AI/ML; Program Synthesis and Machine Learning; Security and Privacy; Logic and automata theory;

Outreach Interests

ConTraIL project: Privacy-preserving contact tracing project. See Whitepaper.

Current research projects and philosophy

My general research philosophy is to turn the theory lens (mainly logic and algorithms) towards practical computing domains. I shy away, for the most part, from working on problems purely motivated by theoretical concerns as well as those entirely based on heuristics to solve practical problems.

My current research interests are primarily in automating software verification, building correct-by-design secure distributed systems, and exploring synergies between machine learning and program synthesis for building intelligent systems.

Automating software verification: Proving a program safe using logical techniques is composed of two phases: finding possible inductive invariants and formally establishing that they are inductive invariants that prove safety.

For the latter problem, the main approach that I currently believe in for automation is the framework of natural proofs that we have been working on for the last 10 years. In particular, though I believe decidable logics are useful for this problem, I do not believe they will solely be powerful enough for addressing the needs of program verification, due to their inherent inexpressivity. Natural proofs aim for automated sound and oftentimes complete (though not decidable) approaches for this problem.

As for the former problem for finding invariants, I believe that a combination of several techniques that are proposed in recent literature (including interpolation and IC3) are called for. The primary techniques techniques I work on have been those based on learning.

Building verified systems: My view on building verified systems is that they should be built correct by construction (as opposed to verifying existing systems), by proving more abstract systems correct and then refining them to concrete performant code. The proofs of abstract systems against complex specifications would invariably require considerable human input and intuition, while automation can help in refining abstract systems to performant code.

Machine Learning and Program Synthesis: One of my latest interests is in building synergies of program synthesis and machine learning--- I believe practical inductive techniques from large data (like neural net models) can be combined with symbolic techniques (learning/searching interpretable concepts, including logical formulas and programs) to build more intelligent systems (see IJCAI paper).

One new permeating theme in my explorations is learning logics. We use learning of logic formulas in applications ranging from software verification (ICE learning), software specification mining (learning preconditions and contracts), and in AI application (interpretable visual discriminators).

Note to prospective students

I am actively looking for Ph.D. students in the following areas. If you are interested, please apply for the Ph.D. program and write to me personally.

Machine learning and program synthesis; learning logics. Exploring new techniques that go beyond neural nets to learn interpretable (symbolic) concepts, abstraction, and reasoning (see IJCAI 2022 and POPL 2022 papers).

Learning logics: Theoretical and practical algorithms for learning logical concepts (see POPL 2022).

Invariant synthesis and mining specifications from code using machine learning, online learning, and learning logical formulas. See papers in CAV14, POPL16, TACAS16, TOCL18, TACAS18, OOPSLA18, and new paper on inductive lemma synthesis.

Logics and reasoning for program verification: especially logics involving expressive specifications for programs manipulating quantified specifications and datastructures, in particular first order logic with least fixpoints. See papers on natural proofs in POPL 2012, PLDI 2013, PLDI 2014, and POPL 2018, work on Frame Logic (ESOP 2020), and new paper on inductive lemma synthesis for FO+lfp.

Building verified software using almost-automated verification, especially distributed systems that involve crptography, like blockchains, bitcoin, and smart contracts (see FMCAD 2019). In particular, I am interested in theory and practice of refining abstract models to concrete distributed models.

Earlier research Projects

Decidable logics for reasoning with heaps/dynamic data structures (Strand)

Security of web browser extensions (VEX)

Testing concurrent programs (Penelope)

Security of access control (VAC)

Annotation and proof mechanisms for safe concurrency (Accord)

Model-checking abstractions of programs (Getafix)

Decidable automata models (See page on visibly pushdown automata)

Software/tools:

VCDryad - An extension of VCC that provides sound but incomplete automatic mechanisms against Dryad specifications, a dialect of separation logic.

Dryad - Sound but incomplete mechanism for proving heap-based programs correct

CAV 2013 tool - for learning universally quantified invariants on linear data-structures using quantified data automata

Strand - Decidable logics/SMT solver for reasoning with heaps

Getafix - A boolean model-checker for concurrent and recursive programs

VEX - Static analysis of web-browser extensions for security vulnerabilities

Penelope - A testing tool for concurrent programs

VAC - Verifier of Access Control

JIST - Java Interface Synthesis Tool

Students and Post-doctoral researchers:

Students currently advised:

Angello Astorga (mining program specifications), jointly advised with Tao Xie

Adithya Murali (combining machine learning and program synthesis; learning logics)

Paul Krogmeier (learning logics)

Shambwaditya Saha (synthesis, learning, mining specifications)

Formerly advised:

Students:

Pranav Garg (deductive verification; compositional verification)

Edgar Pek (static analysis; points-to analysis)

Francesco Sorrentino (testing concurrent programs)

Xiaokang Qiu
Thesis: Automatic Techniques for Proving Correctness of Heap-Manipulating Programs
Now postdoc at MIT/UMD.

Sruthi Bandhakavi; security of web extensions -- now at Google

Postdocs:

Gennaro Parlato -- now faculty at University of Southampton

Karl Palmskog-- now a postdoc at UT Austin.

Daniel Neider (automated deductiver verification, synthesis, learning, invariant synthesis)
Post-doctoral researcher in the ExCAPE project, jointly at Illinois and UCLA. -- now at Max Planck Institute