-
Notifications
You must be signed in to change notification settings - Fork 197
/
advanced-real-numbers.cpp
898 lines (781 loc) · 37.1 KB
/
advanced-real-numbers.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
//==================================================================================
// BSD 2-Clause License
//
// Copyright (c) 2014-2022, NJIT, Duality Technologies Inc. and other contributors
//
// All rights reserved.
//
// Author TPOC: [email protected]
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//==================================================================================
/*
Advanced examples CKKS
*/
// Define PROFILE to enable TIC-TOC timing measurements
#define PROFILE
#include "openfhe.h"
using namespace lbcrypto;
void AutomaticRescaleDemo(ScalingTechnique scalTech);
void ManualRescaleDemo(ScalingTechnique scalTech);
void HybridKeySwitchingDemo1();
void HybridKeySwitchingDemo2();
void FastRotationsDemo1();
void FastRotationsDemo2();
int main(int argc, char* argv[]) {
/*
* Our implementation of CKKS includes four rescaling methods called
* "FIXEDMANUAL*, *FIXEDAUTO*, "FLEXIBLEAUTO", and "FLEXIBLEAUTOEXT".
* THese rescaling methods are explained in the CKKS section of
* https://eprint.iacr.org/2022/915.
*
* Before we start, we need to say a few words about the rescale
* operation, which is central in CKKS. Whenever we multiply two
* ciphertexts c1 and c2 which encrypt numbers m1*D and m2*D
* respectively, we get a result that looks like m1*m2*D^2. Since the
* scaling factor of this number is D^2, we say that the result is of
* depth 2. It is clear that a ciphertext of depth 2 cannot be added
* to ciphertexts of depth 1, because their scaling factors are
* different. Rescaling takes a ciphertext of depth 2, and makes it of
* depth 1 by an operation that looks a lot like dividing by D=2^p.
*
* For efficiency reasons, our implementation of CKKS works in the
* RNS space, which means that we avoid working with big numbers and
* we only work with native integers. One complication that arises
* from this is that we can only rescale by dividing by certain prime
* numbers and not D=2^p.
*
* There are two ways to deal with this. The first is to choose prime
* numbers as close to 2^p as possible, and assume that the scaling
* factor remains the same. This inevitably incurs some approximation
* error, and there are two variants for this scenario: FLEXIBLEAUTO
* and FLEXIBLEAUTOEXT.
*
* The second way of dealing with this is to track how the scaling
* factor changes and try to adjust for it. This is what we do for the
* FLEXIBLEAUTO and FLEXIBALEAUTOEXT variants of CKKS. The tradeoff is
* that FLEXIBLEAUTO* * computations are typically somewhat slower (based on our experience
* the slowdown is around 5-35% depending on the complexity of the
* computation), because of the adjustment of values that need to
* take place.
*
* We have designed FLEXIBLEAUTO(EXT) so it hides all the nuances of
* tracking the depth of ciphertexts and having to call the rescale
* operation. Therefore, FLEXIBLEAUTO(EXT) is more appropriate for users
* who do not want to get into the details of the underlying crypto
* and math, or who want to put together a quick prototype. On the
* contrary, FIXEDMANUAL is more appropriate for production
* applications that have been optimized by experts.
*
* The first two parts of this demo implement the same computation, i.e, the function
* f(x) = x^18 + x^9 + 1, using all four methods.
*
*/
AutomaticRescaleDemo(FLEXIBLEAUTO);
// default
AutomaticRescaleDemo(FLEXIBLEAUTOEXT);
AutomaticRescaleDemo(FIXEDAUTO);
ManualRescaleDemo(FIXEDMANUAL);
/*
* Our implementation of CKKS supports two different algorithms
* for key switching, namely BV and HYBRID. BV corresponds to
* a technique also known as digit decomposition (both RNS and based
* on a digit size). GHS (not implemented separately anymore) corresponds to ciphertext
* modulus doubling, and HYBRID combines the characteristics of both
* BV and GHS. Please refer to the documentation of KeySwitchGen in
* keyswitch-bv.h/cpp and keyswitch-hybrid.h/cpp for more
* details about the different key switch techniques.
*
* For most cases, HYBRID will be the most appropriate and efficient
* key switching technique, and this is why we devote the third and
* fourth part of this demo to HYBRID key switching.
*/
HybridKeySwitchingDemo1();
HybridKeySwitchingDemo2();
/*
* The final parts of this demo showcase our implementation of an
* optimization technique called hoisting. The idea is simple - when
* we want to perform multiple different rotations to the same
* ciphertext, we can compute one part of the rotation algorithm once,
* and reuse it multiple times. Please refer to the documentation of
* EvalFastRotationPrecompute in keyswitch-bv.h/cpp and keyswitch-hybrid.h/cpp
* for more details on hoisting in BV and HYBRID key switching.
*/
FastRotationsDemo1();
FastRotationsDemo2();
return 0;
}
void AutomaticRescaleDemo(ScalingTechnique scalTech) {
/* Please read comments in main() for an introduction to what the
* rescale operation is. Knowing about Rescale() is not necessary
* to use the FLEXIBLEAUTO CKKS variant, it is however needed to
* understand what's happening underneath.
*
* FLEXIBLEAUTO is a variant of CKKS that has two main features:
* 1 - It automatically performs rescaling before every multiplication.
* This is done to make it easier for users to write FHE
* computations without worrying about the depth of ciphertexts
* or rescaling.
* 2 - It tracks the exact scaling factor of all ciphertexts.
* This means that computations in FLEXIBLEAUTO will be more
* accurate than the same computations in FIXEDMANUAL. Keep
* in mind that this difference only becomes apparent when
* dealing with computations of large multiplicative depth; this
* is because a large multiplicative depth means we need to find
* more prime numbers sufficiently close to D=2^p, and this
* becomes harder and harder as the multiplicative depth
* increases.
*/
if (scalTech == FLEXIBLEAUTO) {
std::cout << std::endl << std::endl << std::endl << " ===== FlexibleAutoDemo ============= " << std::endl;
}
else if (scalTech == FLEXIBLEAUTOEXT) {
std::cout << std::endl << std::endl << std::endl << " ===== FlexibleAutoExtDemo ============= " << std::endl;
}
else {
std::cout << std::endl << std::endl << std::endl << " ===== FixedAutoDemo ============= " << std::endl;
}
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(5);
parameters.SetScalingModSize(50);
parameters.SetScalingTechnique(scalTech);
parameters.SetBatchSize(batchSize);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
std::cout << "CKKS scheme is using ring dimension " << cc->GetRingDimension() << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalMultKeyGen(keys.secretKey);
// Input
std::vector<double> x = {1.0, 1.01, 1.02, 1.03, 1.04, 1.05, 1.06, 1.07};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(ptxt, keys.publicKey);
/* Computing f(x) = x^18 + x^9 + 1
*
* In the following we compute f(x) with a computation
* that has a multiplicative depth of 5.
*
* The result is correct, even though there is no call to
* the Rescale() operation.
*/
auto c2 = cc->EvalMult(c, c); // x^2
auto c4 = cc->EvalMult(c2, c2); // x^4
auto c8 = cc->EvalMult(c4, c4); // x^8
auto c16 = cc->EvalMult(c8, c8); // x^16
auto c9 = cc->EvalMult(c8, c); // x^9
auto c18 = cc->EvalMult(c16, c2); // x^18
auto cRes = cc->EvalAdd(cc->EvalAdd(c18, c9), 1.0); // Final result
Plaintext result;
std::cout.precision(8);
cc->Decrypt(cRes, keys.secretKey, &result);
result->SetLength(batchSize);
std::cout << "x^18 + x^9 + 1 = " << result << std::endl;
/*
* Users interested in how FLEXIBLEAUTO works under the
* hood, are welcome to change the line below to "#if 1"
* and observe the changes in scaling factors and depths.
*/
// #if 0
// const auto cryptoParamsCKKS =
// std::dynamic_pointer_cast<CryptoParametersCKKSRNS>(
// cc->GetCryptoParameters());
// std::cout << std::fixed;
// std::cout.precision(2);
// // We have designed FLEXIBLEAUTO so that ciphertexts of a
// // given level have a specific scaling factor.
// std::cout << "\nScaling factors of levels: " << std::endl;
// for (uint32_t i = 0; i < parameters.GetMultiplicativeDepth(); i++) {
// std::cout << "Level " << i << ": "
// << cryptoParamsCKKS->GetScalingFactorReal(i) << std::endl;
// }
// std::cout << std::endl;
// // The initial ciphertext has level 0 and depth 1.
// // One can use additional arguments in
// // cc->MakeCKKSPackedPlaintext(x, depth, level) create
// // plaintexts/ciphertexts at any desired depth/level.
// // Also, in all of the following, notice that scaling
// // factors change at every level and they match the
// // ones printed above.
// std::cout << "Ciphertext c:" << std::endl;
// std::cout << "\t scaling factor: " << c->GetScalingFactor() << std::endl;
// std::cout << "\t scaling factor degree: " << c->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c->GetLevel() << std::endl;
// // Notice how the result of EvalMult(c,c) is of depth 2.
// // This is because the Rescale operation is performed lazily,
// // i.e., only when we want to multiply ciphertexts of depth > 1.
// // Level is still 0 because no rescale operation has been run.
// std::cout << "Ciphertext c2:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c2->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c2->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c2->GetLevel() << std::endl;
// // Here, the c2 inputs where of depth 2, so they had to be rescaled
// // before computing c4. Hence, the level has become 1. Again,
// // rescaling happens lazily, so depth of the result is 2.
// std::cout << "Ciphertext c4:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c4->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c4->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c4->GetLevel() << std::endl;
// std::cout << "Ciphertext c8:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c8->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c8->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c8->GetLevel() << std::endl;
// std::cout << "Ciphertext c16:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c16->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c16->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c16->GetLevel() << std::endl;
// // c9 is the result of EvalMult between c8 (depth:2, level:2) and
// // c (depth: 1, level: 0). Inputs are automatically brought to the
// // correct depth and level before the operation and the user does
// // not need to care about these low level details.
// std::cout << "Ciphertext c9:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c9->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c9->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c9->GetLevel() << std::endl;
// std::cout << "Ciphertext c18:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(c18->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << c18->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << c18->GetLevel() << std::endl;
// // The result has the same depth and level as c18 because
// // additions do not require any rescaling.
// std::cout << "Ciphertext cRes:" << std::endl;
// std::cout << "\t scaling factor: (" << sqrt(cRes->GetScalingFactor()) << ")^2"
// << std::endl;
// std::cout << "\t scaling factor degree: " << cRes->GetNoiseScaleDeg() << std::endl;
// std::cout << "\t level: " << cRes->GetLevel() << std::endl;
// std::cout << std::defaultfloat;
// std::cout.precision(8);
// #endif
}
void ManualRescaleDemo(ScalingTechnique scalTech) {
/* Please read comments in main() for an introduction to what the
* rescale operation is, and what's the FIXEDMANUAL variant of CKKS.
*
* Even though FIXEDMANUAL does not implement automatic rescaling
* as FLEXIBLEAUTO does, this does not mean that it does not abstract
* away some of the nitty-gritty details of using CKKS.
*
* In CKKS, ciphertexts are defined versus a large ciphertext modulus Q.
* Whenever we rescale a ciphertext, its ciphertext modulus becomes
* smaller too. All homomorphic operations require that their inputs are
* defined over the same ciphertext modulus, and therefore, we need to
* adjust one of them if their ciphertext moduli do not match. The way
* this is done in the original CKKS paper is through an operation called
* Modulus Switch. In our implementation, we call this operation
* LevelReduce, and both FIXEDMANUAL and FLEXIBLEAUTO do it automatically.
* As far as we know, automatic level reduce does not incur any performance
* penalty and this is why it is performed in both FIXEDMANUAL and
* FLEXIBLEAUTO.
*
* Overall, we believe that automatic modulus switching and rescaling make
* CKKS much easier to use, at least for non-expert users.
*/
std::cout << "\n\n\n ===== FixedManualDemo ============= " << std::endl;
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(5);
parameters.SetScalingModSize(50);
parameters.SetBatchSize(batchSize);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
std::cout << "CKKS scheme is using ring dimension " << cc->GetRingDimension() << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalMultKeyGen(keys.secretKey);
// Input
std::vector<double> x = {1.0, 1.01, 1.02, 1.03, 1.04, 1.05, 1.06, 1.07};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(keys.publicKey, ptxt);
/* Computing f(x) = x^18 + x^9 + 1
*
* Compare the following with the corresponding code
* for FLEXIBLEAUTO. Here we need to track the depth of ciphertexts
* and call Rescale() whenever needed. In this instance it's still
* not hard to do so, but this can be quite tedious in other
* complicated computations (e.g., in bootstrapping).
*
*/
// x^2
auto c2_depth2 = cc->EvalMult(c, c);
auto c2_depth1 = cc->Rescale(c2_depth2);
// x^4
auto c4_depth2 = cc->EvalMult(c2_depth1, c2_depth1);
auto c4_depth1 = cc->Rescale(c4_depth2);
// x^8
auto c8_depth2 = cc->EvalMult(c4_depth1, c4_depth1);
auto c8_depth1 = cc->Rescale(c8_depth2);
// x^16
auto c16_depth2 = cc->EvalMult(c8_depth1, c8_depth1);
auto c16_depth1 = cc->Rescale(c16_depth2);
// x^9
auto c9_depth2 = cc->EvalMult(c8_depth1, c);
// x^18
auto c18_depth2 = cc->EvalMult(c16_depth1, c2_depth1);
// Final result
auto cRes_depth2 = cc->EvalAdd(cc->EvalAdd(c18_depth2, c9_depth2), 1.0);
auto cRes_depth1 = cc->Rescale(cRes_depth2);
Plaintext result;
std::cout.precision(8);
cc->Decrypt(keys.secretKey, cRes_depth1, &result);
result->SetLength(batchSize);
std::cout << "x^18 + x^9 + 1 = " << result << std::endl;
}
void HybridKeySwitchingDemo1() {
/*
* Please refer to comments in the demo-simple_real_number.cpp
* for a brief introduction on what key switching is and to
* find reference for HYBRID key switching.
*
* In this demo, we focus on how to choose the number of digits
* in HYBRID key switching, and how that affects the usage and
* efficiency of the CKKS scheme.
*
*/
std::cout << "\n\n\n ===== HybridKeySwitchingDemo1 ============= " << std::endl;
/*
* dnum is the number of large digits in HYBRID decomposition
*
* If not supplied (or value 0 is supplied), the default value is
* set as follows:
* - If multiplicative depth is > 3, then dnum = 3 digits are used.
* - If multiplicative depth is 3, then dnum = 2 digits are used.
* - If multiplicative depth is < 3, then dnum is set to be equal to
* multDepth+1
*/
uint32_t dnum = 2;
/* To understand the effects of changing dnum, it is important to
* understand how the ciphertext modulus size changes during key
* switching.
*
* In our RNS implementation of CKKS, every ciphertext corresponds
* to a large number (which is represented as small integers in RNS)
* modulo a ciphertext modulus Q, which is defined as the product of
* (multDepth+1) prime numbers: Q = q0 * q1 * ... * qL. Each qi is
* selected to be close to the scaling factor D=2^p, hence the total
* size of Q is approximately:
*
* sizeof(Q) = (multDepth+1)*scaleModSize.
*
* HYBRID key switching takes a number d that's defined modulo Q,
* and performs 4 steps:
* 1 - Digit decomposition:
* Split d into dnum digits - the size of each digit is roughly
* ceil(sizeof(Q)/dnum)
* 2 - Extend ciphertext modulus from Q to Q*P
* Here P is a product of special primes
* 3 - Multiply extended component with key switching key
* 4 - Decrease the ciphertext modulus back down to Q
*
* It's not necessary to understand how all these stages work, as
* long as it's clear that the size of the ciphertext modulus is
* increased from sizeof(Q) to sizeof(Q)+sizeof(P) in stage 2. P
* is always set to be as small as possible, as long as sizeof(P)
* is larger than the size of the largest digit, i.e., than
* ceil(sizeof(Q)/dnum). Therefore, the size of P is inversely
* related to the number of digits, so the more digits we have, the
* smaller P has to be.
*
* The tradeoff here is that more digits means that the digit
* decomposition stage becomes more expensive, but the maximum
* size of the ciphertext modulus Q*P becomes smaller. Since
* the size of Q*P determines the necessary ring dimension to
* achieve a certain security level, more digits can in some
* cases mean that we can use smaller ring dimension and get
* better performance overall.
*
* We show this effect with demos HybridKeySwitchingDemo1 and
* HybridKeySwitchingDemo2.
*
*/
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(5);
parameters.SetScalingModSize(50);
parameters.SetBatchSize(batchSize);
parameters.SetScalingTechnique(FLEXIBLEAUTO);
parameters.SetNumLargeDigits(dnum);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
std::cout << "CKKS scheme is using ring dimension " << cc->GetRingDimension() << std::endl;
std::cout << "- Using HYBRID key switching with " << dnum << " digits" << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalRotateKeyGen(keys.secretKey, {1, -2});
// Input
std::vector<double> x = {1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(keys.publicKey, ptxt);
TimeVar t;
TIC(t);
auto cRot1 = cc->EvalRotate(c, 1);
auto cRot2 = cc->EvalRotate(cRot1, -2);
double time2digits = TOC(t);
// Take note and compare the runtime to the runtime
// of the same computation in the next demo.
Plaintext result;
std::cout.precision(8);
cc->Decrypt(keys.secretKey, cRot2, &result);
result->SetLength(batchSize);
std::cout << "x rotate by -1 = " << result << std::endl;
std::cout << " - 2 rotations with HYBRID (2 digits) took " << time2digits << "ms" << std::endl;
/* Interested users may set the following if to 1
* to observe the prime numbers comprising Q and P,
* and how these change with the number of digits
* dnum.
*/
// #if 0
// const auto cryptoParamsCKKS =
// std::dynamic_pointer_cast<CryptoParametersCKKSRNS>(
// cc->GetCryptoParameters());
// auto paramsQ = cc->GetElementParams()->GetParams();
// std::cout << "\nModuli in Q:" << std::endl;
// for (uint32_t i = 0; i < paramsQ.size(); i++) {
// // q0 is a bit larger because its default size is 60 bits.
// // One can change this by supplying the firstModSize argument
// // in genCryptoContextCKKS.
// std::cout << "q" << i << ": " << paramsQ[i]->GetModulus() << std::endl;
// }
// auto paramsQP = cryptoParamsCKKS->GetParamsQP();
// std::cout << "Moduli in P: " << std::endl;
// BigInteger P = BigInteger(1);
// for (uint32_t i = 0; i < paramsQP->GetParams().size(); i++) {
// if (i > paramsQ.size()) {
// P = P * BigInteger(paramsQP->GetParams()[i]->GetModulus());
// std::cout << "p" << i - paramsQ.size() << ": "
// << paramsQP->GetParams()[i]->GetModulus() << std::endl;
// }
// }
// auto QBitLength = cc->GetModulus().GetLengthForBase(2);
// auto PBitLength = P.GetLengthForBase(2);
// std::cout << "\nQ = " << cc->GetModulus() << " (bit length: " << QBitLength
// << ")" << std::endl;
// std::cout << "P = " << P << " (bit length: " << PBitLength << ")"
// << std::endl;
// std::cout << "Total bit-length of ciphertext modulus: "
// << QBitLength + PBitLength << std::endl;
// std::cout << "Given this ciphertext modulus, a ring dimension of "
// << cc->GetRingDimension() << " gives us 128-bit security."
// << std::endl;
// #endif
}
void HybridKeySwitchingDemo2() {
/*
* Please refer to comments in HybridKeySwitchingDemo1.
*
*/
std::cout << "\n\n\n ===== HybridKeySwitchingDemo2 ============= " << std::endl;
/*
* Here we use dnum = 3 digits. Even though 3 digits are
* more than the two digits in the previous demo and the
* cost of digit decomposition is higher, the increase in
* digits means that individual digits are smaller, and we
* can perform key switching by using only one special
* prime in P (instead of two in the previous demo).
*
* This also means that the maximum size of ciphertext
* modulus in key switching is smaller by 60 bits, and it
* turns out that this decrease is adequate to warrant a
* smaller ring dimension to achieve the same security
* level (128-bits).
*
*/
uint32_t dnum = 3;
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(5);
parameters.SetScalingModSize(50);
parameters.SetBatchSize(batchSize);
parameters.SetScalingTechnique(FLEXIBLEAUTO);
parameters.SetNumLargeDigits(dnum);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
// Compare the ring dimension in this demo to the one in
// the previous.
std::cout << "CKKS scheme is using ring dimension " << cc->GetRingDimension() << std::endl;
std::cout << "- Using HYBRID key switching with " << dnum << " digits" << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalRotateKeyGen(keys.secretKey, {1, -2});
// Input
std::vector<double> x = {1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(keys.publicKey, ptxt);
TimeVar t;
TIC(t);
auto cRot1 = cc->EvalRotate(c, 1);
auto cRot2 = cc->EvalRotate(cRot1, -2);
// The runtime here is smaller than in the previous demo.
double time3digits = TOC(t);
Plaintext result;
std::cout.precision(8);
cc->Decrypt(keys.secretKey, cRot2, &result);
result->SetLength(batchSize);
std::cout << "x rotate by -1 = " << result << std::endl;
std::cout << " - 2 rotations with HYBRID (3 digits) took " << time3digits << "ms" << std::endl;
/* Interested users may set the following if to 1
* to observe the prime numbers comprising Q and P,
* and how these change with the number of digits
* dnum.
*/
// #if 0
// const auto cryptoParamsCKKS =
// std::dynamic_pointer_cast<CryptoParametersCKKSRNS>(
// cc->GetCryptoParameters());
// auto paramsQ = cc->GetElementParams()->GetParams();
// std::cout << "\nModuli in Q:" << std::endl;
// for (uint32_t i = 0; i < paramsQ.size(); i++) {
// // q0 is a bit larger because its default size is 60 bits.
// // One can change this by supplying the firstModSize argument
// // in genCryptoContextCKKS.
// std::cout << "q" << i << ": " << paramsQ[i]->GetModulus() << std::endl;
// }
// auto paramsQP = cryptoParamsCKKS->GetParamsQP();
// std::cout << "Moduli in P: " << std::endl;
// BigInteger P = BigInteger(1);
// for (uint32_t i = 0; i < paramsQP->GetParams().size(); i++) {
// if (i > paramsQ.size()) {
// P = P * BigInteger(paramsQP->GetParams()[i]->GetModulus());
// std::cout << "p" << i - paramsQ.size() << ": "
// << paramsQP->GetParams()[i]->GetModulus() << std::endl;
// }
// }
// auto QBitLength = cc->GetModulus().GetLengthForBase(2);
// auto PBitLength = P.GetLengthForBase(2);
// std::cout << "\nQ = " << cc->GetModulus() << " (bit length: " << QBitLength
// << ")" << std::endl;
// std::cout << "P = " << P << " (bit length: " << PBitLength << ")"
// << std::endl;
// std::cout << "Given this ciphertext modulus, a ring dimension of "
// << cc->GetRingDimension() << " gives us 128-bit security."
// << std::endl;
// #endif
}
void FastRotationsDemo1() {
/*
* In CKKS, whenever someone applies a rotation R() to a ciphertext
* encrypted with key s, we get a result which is not valid under
* key s, but under the same rotation R(s) of s. Therefore, after
* every rotation we need to perform key switching, making them as
* expensive as multiplications.
*
* As mentioned earlier (in comments of HybridKeySwitchingDemo1),
* key switching involves the following steps:
* 1 - Digit decomposition
* 2 - Extend ciphertext modulus from Q to Q*P
* 3 - Multiply extended component with key switching key
* 4 - Decrease the ciphertext modulus back down to Q
*
* A useful observation is that the first two steps are independent
* of the particular rotation we want to perform. Steps 3-4 on the
* other hand depend on the specific rotation we have at hand,
* because each rotation index has a different key switch key.
*
* This observation means that, if we want to perform multiple
* different rotations to the same ciphertext, we can perform
* the first two steps once, and then only perform steps 3-4 for
* each rotation. This technique is called hoisting, and we have
* implemented it for all three key switching techniques (BV, GHS,
* HYBRID) in OpenFHE.
*
* The benefits expected by this technique differ depending on the
* key switching algorithms we're using. BV is the technique that
* gets the greatest benefits, because the digit decomposition is
* the most expensive part. However, HYBRID also benefits from
* hoisting, and we show this in this part of the demo.
*
*/
std::cout << "\n\n\n ===== FastRotationsDemo1 ============= " << std::endl;
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(1);
parameters.SetScalingModSize(50);
parameters.SetBatchSize(batchSize);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
uint32_t N = cc->GetRingDimension();
std::cout << "CKKS scheme is using ring dimension " << N << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalRotateKeyGen(keys.secretKey, {1, 2, 3, 4, 5, 6, 7});
// Input
std::vector<double> x = {0, 0, 0, 0, 0, 0, 0, 1};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(keys.publicKey, ptxt);
Ciphertext<DCRTPoly> cRot1, cRot2, cRot3, cRot4, cRot5, cRot6, cRot7;
// First, we perform 7 regular (non-hoisted) rotations
// and measure the runtime.
TimeVar t;
TIC(t);
cRot1 = cc->EvalRotate(c, 1);
cRot2 = cc->EvalRotate(c, 2);
cRot3 = cc->EvalRotate(c, 3);
cRot4 = cc->EvalRotate(c, 4);
cRot5 = cc->EvalRotate(c, 5);
cRot6 = cc->EvalRotate(c, 6);
cRot7 = cc->EvalRotate(c, 7);
double timeNoHoisting = TOC(t);
auto cResNoHoist = c + cRot1 + cRot2 + cRot3 + cRot4 + cRot5 + cRot6 + cRot7;
// M is the cyclotomic order and we need it to call EvalFastRotation
uint32_t M = 2 * N;
// Then, we perform 7 rotations with hoisting.
TIC(t);
auto cPrecomp = cc->EvalFastRotationPrecompute(c);
cRot1 = cc->EvalFastRotation(c, 1, M, cPrecomp);
cRot2 = cc->EvalFastRotation(c, 2, M, cPrecomp);
cRot3 = cc->EvalFastRotation(c, 3, M, cPrecomp);
cRot4 = cc->EvalFastRotation(c, 4, M, cPrecomp);
cRot5 = cc->EvalFastRotation(c, 5, M, cPrecomp);
cRot6 = cc->EvalFastRotation(c, 6, M, cPrecomp);
cRot7 = cc->EvalFastRotation(c, 7, M, cPrecomp);
double timeHoisting = TOC(t);
// The time with hoisting should be faster than without hoisting.
auto cResHoist = c + cRot1 + cRot2 + cRot3 + cRot4 + cRot5 + cRot6 + cRot7;
Plaintext result;
std::cout.precision(8);
cc->Decrypt(keys.secretKey, cResNoHoist, &result);
result->SetLength(batchSize);
std::cout << "Result without hoisting = " << result << std::endl;
std::cout << " - 7 rotations on x without hoisting took " << timeNoHoisting << "ms" << std::endl;
cc->Decrypt(keys.secretKey, cResHoist, &result);
result->SetLength(batchSize);
std::cout << "Result with hoisting = " << result << std::endl;
std::cout << " - 7 rotations on x with hoisting took " << timeHoisting << "ms" << std::endl;
}
void FastRotationsDemo2() {
/*
* This demo is identical to the previous one, with the exception
* that we use BV key switching instead of HYBRID.
*
* The benefits expected by hoisting differ depending on the
* key switching algorithms we're using. BV is the technique that
* gets the greatest benefits, because the digit decomposition is
* the most expensive part. However, HYBRID also benefits from
* hoisting, and we show this in this part of the demo.
*
*/
std::cout << "\n\n\n ===== FastRotationsDemo2 ============= " << std::endl;
// uint32_t dnum = 0; -already default
/*
* This controls how many multiplications are possible without rescaling.
* The number of multiplications (maxRelinSkDeg) is maxDepth - 1.
* This is useful for an optimization technique called lazy
* re-linearization (only applicable in FIXEDMANUAL, as
* FLEXIBLEAUTO implements automatic rescaling).
*/
// uint32_t maxDepth (maxRelinSkDeg) = 2; - already default
/*
* The digit size is only used in BV key switching and
* it allows us to perform digit decomposition at a finer granularity.
* Under normal circumstances, digit decomposition is what we call
* RNS decomposition, i.e., each digit is roughly the size of the
* qi's that comprise the ciphertext modulus Q. When using BV, in
* certain cases like having to perform rotations without any
* preceding multiplication, we need to have smaller digits to prevent
* noise from corrupting the result. In this case, using digitSize = 10
* does the trick. Users are encouraged to set this to 0 (i.e., RNS
* decomposition) and see how the results are incorrect.
*/
uint32_t digitSize = 10;
uint32_t batchSize = 8;
CCParams<CryptoContextCKKSRNS> parameters;
parameters.SetMultiplicativeDepth(1);
parameters.SetScalingModSize(50);
parameters.SetBatchSize(batchSize);
parameters.SetScalingTechnique(FLEXIBLEAUTO);
parameters.SetKeySwitchTechnique(BV);
parameters.SetFirstModSize(60);
parameters.SetDigitSize(digitSize);
CryptoContext<DCRTPoly> cc = GenCryptoContext(parameters);
uint32_t N = cc->GetRingDimension();
std::cout << "CKKS scheme is using ring dimension " << N << std::endl << std::endl;
cc->Enable(PKE);
cc->Enable(KEYSWITCH);
cc->Enable(LEVELEDSHE);
auto keys = cc->KeyGen();
cc->EvalRotateKeyGen(keys.secretKey, {1, 2, 3, 4, 5, 6, 7});
// Input
std::vector<double> x = {0, 0, 0, 0, 0, 0, 0, 1};
Plaintext ptxt = cc->MakeCKKSPackedPlaintext(x);
std::cout << "Input x: " << ptxt << std::endl;
auto c = cc->Encrypt(keys.publicKey, ptxt);
Ciphertext<DCRTPoly> cRot1, cRot2, cRot3, cRot4, cRot5, cRot6, cRot7;
// First, we perform 7 regular (non-hoisted) rotations
// and measure the runtime.
TimeVar t;
TIC(t);
cRot1 = cc->EvalRotate(c, 1);
cRot2 = cc->EvalRotate(c, 2);
cRot3 = cc->EvalRotate(c, 3);
cRot4 = cc->EvalRotate(c, 4);
cRot5 = cc->EvalRotate(c, 5);
cRot6 = cc->EvalRotate(c, 6);
cRot7 = cc->EvalRotate(c, 7);
double timeNoHoisting = TOC(t);
auto cResNoHoist = c + cRot1 + cRot2 + cRot3 + cRot4 + cRot5 + cRot6 + cRot7;
// M is the cyclotomic order and we need it to call EvalFastRotation
uint32_t M = 2 * N;
// Then, we perform 7 rotations with hoisting.
TIC(t);
auto cPrecomp = cc->EvalFastRotationPrecompute(c);
cRot1 = cc->EvalFastRotation(c, 1, M, cPrecomp);
cRot2 = cc->EvalFastRotation(c, 2, M, cPrecomp);
cRot3 = cc->EvalFastRotation(c, 3, M, cPrecomp);
cRot4 = cc->EvalFastRotation(c, 4, M, cPrecomp);
cRot5 = cc->EvalFastRotation(c, 5, M, cPrecomp);
cRot6 = cc->EvalFastRotation(c, 6, M, cPrecomp);
cRot7 = cc->EvalFastRotation(c, 7, M, cPrecomp);
double timeHoisting = TOC(t);
/* The time with hoisting should be faster than without hoisting.
* Also, the benefits from hoisting should be more pronounced in this
* case because we're using BV. Of course, we also observe less
* accurate results than when using HYBRID, because of using
* digitSize = 10 (Users can decrease digitSize to see the accuracy
* increase, and performance decrease).
*/
auto cResHoist = c + cRot1 + cRot2 + cRot3 + cRot4 + cRot5 + cRot6 + cRot7;
Plaintext result;
std::cout.precision(8);
cc->Decrypt(keys.secretKey, cResNoHoist, &result);
result->SetLength(batchSize);
std::cout << "Result without hoisting = " << result << std::endl;
std::cout << " - 7 rotations on x without hoisting took " << timeNoHoisting << "ms" << std::endl;
cc->Decrypt(keys.secretKey, cResHoist, &result);
result->SetLength(batchSize);
std::cout << "Result with hoisting = " << result << std::endl;
std::cout << " - 7 rotations on x with hoisting took " << timeHoisting << "ms" << std::endl;
}