Confidential machine learning on untrusted platforms: a survey

Cybersecurity

Table 2 Examples for primitive switching strategies in hybrid composition of CML frameworks

Framework	Primitive Switch	Operation Switch	Justification
Sharma and Chen (2019)	SHE → GC	Matrix vector multiplication → Sign Check	Sign checking is impractically expensive with SHE whereas tolerable with GC.
Nikolaenko et al. (2013)	AHE → GC	Matrix Additions → Cholesky’s decomposition	The operations of division and square root in Cholesky’s decomposition were not feasible with the AHE scheme.
Nikolaenko et al. (2013)	AHE → GC	Matrix Additions → Gradient Descent	Gradient descent involved multiplications, additions, and subtractions not entirely feasible with the AHE scheme.
Mohassel and Zhang (2017)	SecSh → GC	Matrix-vector multiplication → Comparison	Comparison is impossible over randomly shared secrets leading the switch to the garbled circuits.
Mohassel and Zhang (2017)	GC → SecSh	Comparison → Vector Subtraction	Use of garbled circuits for comparison was unavoidable however continuing GC on to vector subtraction would result in excessive cost overhead.
Demmler et al. (2015)	SecSh → AHE/OT	Data at rest → Multiplication	Multiplication with random shares required switching to either AHE or OT protocol involving the two parties in the frameworks.
Riazi et al. (2018)	SecSh → GC	Matrix matrix multiplication → ReLu computation	Sign checking is impossible over randomly shared secrets leading the switch to garbled circuits.
Riazi et al. (2018)	GC →SecSh	ReLu → Matrix vector multiplication	Use of garbled circuits for matrix vector multiplication is impractical.