Analysis and Synthesis
Replicated systems are widely deployed to tolerate faults in aviation
control systems and geo-distributed data stores. Given high-level sequential specifications, we automatically analyze
the specifications and synthesize consistent, convergent and coordination-avoiding replicated distributed systems.
See the POPL'19 paper
Certified Key-value Stores
We implement, specify and mechanically verify distributed key-value
stores in proof assistants.
See the POPL'16 paper
Compositional Verification of Distributed System Stacks
What are the fundamental principles for compositional verification of
distributed systems? We design program logics and compositional
verification techniques to build certified distributed system stacks.
See our NSF-supported project.
Ongoing project: Automatic
Synthesis of Parallel and Distributed Graph-processing
Given a high-level specification of a graph processing problem, we
automatically synthesize programs for multiple parallel and distributed
Blockchains and Cryptocurrencies
project: Efficient Blockchain Protocols
Proof-of-work protocol wastes a huge amount
of computational power. What are more
energy-efficient blockchain protocols?
project: Certified Blockchains with Explicit Safety and Security Guarantees
What are the correctness conditions
and verification principles for blockchain
protocols? We aim at building certified blockchains with explicit safety and
security guarantees using proof assistants.
Transaction Specification, Testing and Verification
What are the correctness criteria
and proof techniques for transactions?
the DISC'14 paper, WTTM'12 and WTTM'13 papers.
We presented a testing technique
that found bugs in transaction algorithms.
See the DISC'13 paper.
We have built a mechanically
checked framework for verification of transactional
See the CONCUR'12 paper.
specify and implement
safe transactions in the presence of communication between them.
See the PPoPP'12 paper.
of Concurrent Data
Composing atomic concurrent
operations is challenging and researchers have found a myriad of atomicity bugs in compositions of concurrent
libraries. We automatically verify the
atomicity of compositions.
See the CAV'14 paper.
To gain performance, compilers and processors reorder
program instructions. Reordering instructions may violate
safety of concurrent executions. Experts insert fence
instructions to prevent violating reorders. We automate fence insertion.
See the OOPSLA'15 paper
and PODC'17 paper and DISC'19 paper.
Composable Concurrent Programming
We are building a certified
concurrent collections library that
supports atomic composition of method calls across objects.
Trust and Security
Ongoing project: Memory safety
attackers can use out-of-bound memory
accesses to take over the whole
system. We develop static and dynamic techniques to ensure that memory
accesses are within bounds and within the access capabilities of the enclosing module.
See our NSF-supported project.
for Secure Cloud Computations
To protect user privacy, we
encrypt data before processing it in the
cloud. We apply type inference to find the most efficient encryption
See the OOPSLA'13 paper.
Aggregation and Inference in Social Networks
Given local trust
information between direct friends, we automatically infer a trust
measure between any pair of users.
See the JCI'09 paper
Ongoing Project: Robust Machine Learning
Reliable deployment of machine
learning in mission- and life-critical applications requires the
specification and verification of their safety and security properties.
What are the specification and verification methods for reliability of
learned models? Can we devise training algorithms with reliability
guarantees or verify learned models?
Performance Models to Guide
Synthesis of Efficient Concurrent Data Structures
a high-level relational specification of
data, we want to automatically synthesize efficient concurrent data
structures. Performance is an irregular phenomenon that can be only
learned from experiments. We train neural networks as performance
models. We use the performance model to choose the most efficient
candidate from the space of possible candidate data structures.
between Q-Learning Agents
We presented a
technique for cooperation between heterogeneous Q-learning agents.
See the SMC'04 paper.
Language and Type System for Programming Biochemistry
We introduce the BioScript domain-specific language for programmable
biochemistry which executes on emerging microfluidic platforms and the
ChemType type system that ensures that the chemical interactions are
See the OOPSLA'18 paper
Ongoing project: Drug Synthesis
Can we apply program synthesis techniques to
automatically synthesize drugs with specified properties?