Maintenance

• Fixed security bug: malicious parties not committed to MAC key across threads and executions
• Fixed security bug: insufficient checks in Rep4
• Improved radix sort
• Optimized 32-bit computation
• Multithreaded Rep3 shuffling
• Improved benchmarking in binary virtual machines
• More sophisticated register allocation
• C++-20 requirement throughout
• Option for parametrized communication cost
• Use budget when merging function calls
• Timeout for compilation to allow identification of blowouts
• Option to disable index checks in loops
• File persistence for public integers

Expected communication cost in compiler

• Expected communication cost in compiler
• Semi-honest option of Rep4
• Reduced communication for preprocessing in Dealer protocol
• Option of choosing SoftSpoken parameter at run-time
• BERT functionality (@hiddely)
• Recommended reading list in documentation
• Fixed security bug: minor security parameter degradation in HighGear input protocol

Protocols with function-dependent preprocessing

• Add protocols with function-dependent preprocessing (https://eprint.iacr.org/2025/919)
• Parallelize shuffling (@vincent-ehrmanntraut)
• More efficient probabilistic truncation in Rep3
• More efficient binary to arithmetic conversion for one bit in Rep3
• Backend optimizations benefitting the most efficient protocols like Rep3
• Allow regint registers as argument in exported functions
• More efficient dot product for GF(2^n)
• File persistence for GF(2^n)
• Output of binary secrets
• SHA256
• Improved navigation by providing links to relevant papers (./compile.py --papers) and outputting which code is executed (./<protocol>-party.x --code-locations)
• Fixed security bug: remove MAC key in case of failure

Functionality to call high-level code from C++

• Functionality to call high-level code from C++
• Matrix triples from file for all appropriate protocols
• Exit with message on errors instead of uncaught exceptions
• Reduce memory usage for binary memory
• Optimized cint-regint conversion in Dealer protocol
• Fixed security bug: missing MAC check in probabilistic truncation

Maintenance

• Inference with non-sequential PyTorch networks
• SHA-3 for any input length (@hiddely)
• Improved client facilities
• Shuffling with malicious security for SPDZ-wise protocols by Asharov et al.
• More reusable bytecode via in-thread calling facility
• Recursive functions without return values
• Fewer rounds for parallel matrix multiplications (@vincent-ehrmanntraut)
• Optimized usage of SoftSpokenOT in semi-honest protocols
• More integrity checks on storage in MAC-based protocols
• Use C++17
• Use glibc 2.18 for the binaries
• Fixed security bugs: remotely caused buffer overflows (#1382)
• Fixed security bug: Missing randomization before revealing to client
• Fixed security bug: Bias in Rep3 secure shuffling

Multinode computation

• Functionality for multiple nodes per party
• Functionality to use disk space for high-level data structures
• True division is always fixed-point division (similar to Python 3)
• Compiler option to optimize for specific protocol
• Cleartext permutation
• Faster compilation and lower bytecode size
• Functionality to output secret shares from high-level code
• Run-time command-line arguments accessible from high-level code
• Client connection setup specifies cleartext domain
• Compile-time parameter for connection timeout
• Prevent connections from timing out (@ParallelogramPal)
• More ECDSA examples
• More flexible multiplication instruction
• Dot product instruction supports several operations at once
• Example-based virtual machine explanation

Maintenance

• Path Oblivious Heap (@tskovlund)
• Adjust batch and bucket size to program
• Direct communication available in more protocols
• Option for seed in fake preprocessing (@strieflin)
• Lower memory usage due to improved register allocation
• New instructions to speed up CISC compilation
• Protocol implementation example
• Fixed security bug: missing MAC checks in multi-threaded programs
• Fixed security bug: race condition in MAC check
• Fixed security bug: missing shuffling check in PS mod 2^k and Brain
• Fixed security bug: insufficient drowning in pairwise protocols

Maintenance

• More extensive benchmarking outputs
• Replace MPIR by GMP
• Secure reading of edaBits from files
• Semi-honest client communication
• Back-propagation for average pooling
• Parallelized convolution
• Probabilistic truncation as in ABY3
• More balanced communication in Shamir secret sharing
• Avoid unnecessary communication in Dealer protocol
• Linear solver using Cholesky decomposition
• Accept .py files for compilation
• Fixed security bug: proper accounting for random elements

More accessible machine learning functionality

• Easier-to-use machine learning interface
• Integrated compilation-execution facility
• Import/export sequential models and parameters from/to PyTorch
• Binary-format input files
• Less aggressive round optimization for faster compilation by default
• Multithreading with client interface
• Functionality to protect order of specific memory accesses
• Oblivious transfer works again on older (pre-2011) x86 CPUs
• clang is used by default

Decision tree training

• Decision tree training
• Optimized oblivious shuffle in Rep3
• Optimized daBit generation in Rep3 and semi-honest HE-based 2PC
• Optimized element-vector AND in SemiBin
• Optimized input protocol in Shamir-based protocols
• Square-root ORAM (@Quitlox)
• Improved ORAM in binary circuits
• UTF-8 outputs